OKLMdocumentation v1.0.3
On this page

Command reference

Every command of oklm and every tool of oklm-gui on one page: its synopsis, what it does, what each option means, and the page that describes it in depth. The synopses are the program's own usage text (oklm help, oklm-gui --tool NAME --help), copied here by python3 tools/docs/interfaces.py synopsis; make docs-lint fails when one of them is no longer what the sources print, or when a command has none.

Status: maintained; the synopses are generated from apps/oklm.c and apps/gui/*.c, the rest is written by hand.

Reading a synopsis

  • <name> is a value to give, [...] is optional, a|b is a choice, and ... after an item means it may be repeated.
  • A number written after an option ([-n 128], [--port 7430]) is its default.
  • Exit status: 0 on success, 1 on a failure, 2 on a usage error (the command's synopsis is then printed on the error stream); the commands with other statuses say so below.
  • oklm help (also --help, -h) prints the whole usage text; oklm transcribe --help and oklm speak --help print their own.
  • The environment variables are in build.md.
Area Commands
The program version, devices, types, help
Files info, convert, quantize, imatrix, split, merge, lora-merge, expert-pack, tokenize, train-tokenizer, ingest
Models and memory models, plan, tune
Inference run, chat, ppl, embed, rerank
Training train, rl, abliterate, refusals
Serving serve
Agent agent, permissions, audit
Retrieval rag index, rag search, rag info, rag eval
Multimodal and speech describe, transcribe, speak
Graphical tools oklm-gui --tool launcher, chat, models, servers, inspect, convert, train, sessions, rag

Shared options

These options mean the same thing in every command that takes them. The commands below name the groups they take and describe only what is their own.

The model

Option Meaning
-m MODEL, --model MODEL the model: a GGUF file (any shard of a split model opens the whole model), a Hugging Face checkpoint directory where the synopsis says hf_dir, or the name of an installed catalogue model (models.md)
--mmproj FILE a projector (vision or audio encoder) beside the model; a catalogue name works too, and a catalogue model that carries a projector gets it without the option where the command says so
--lora ADAPTER, --lora-scale S a LoRA adapter merged into the model as it loads, times S (1) (run, chat, ppl, serve; training.md)
--chat-template FILE|NAME the chat template used in place of the file's: a Jinja file, or a built-in format's name (vicuna, chatml, llama2, llama3, mistral, mistral-v3, gemma, phi3, alpaca, deepseek-coder, foundation-sec) (run, chat, serve, agent, describe; inference.md)
--template-vars JSON an object of more variables given to the chat template, such as {"enable_thinking": false}

Everywhere a model file is expected (-m, --model, --mmproj, --draft, --embed-model, --rerank-model, --transcribe-model, --speak-model) the name of an installed catalogue model can be given.

Where it computes

Option Meaning
--backend cpu|opencl|vulkan|auto the compute backend; auto takes a device when there is one. The training commands take auto|cpu|opencl|rocm instead (opencl.md, vulkan.md, rocm_spec.md)
--gpu-layers N, -ngl N on a device backend, the model's last N layers and its output head on the device, the others on the CPU; without it, as many as fit in the device's memory
--gpu-stream N with --gpu-layers below the layer count: batches of N tokens or more stream the other layers to the device (run, chat, ppl; plan counts it)
--devices N the device layers spread over N devices of the backend (0, the default: every device of OKLM_CL_DEVICES / OKLM_VK_DEVICES)
-c CTX the context length in tokens; without it 4096, or the model's trained context when that is smaller (agent: 32768, capped the same way) (models.md)
-ctk TYPE, -ctv TYPE, --cache-type-k, --cache-type-v the element type of the KV cache's keys and values: f16 (default), f32, bf16, q8_0, q4_0, q4_1, q5_0, q5_1, iq4_nl
--tune (run) tune the CPU kernels' work split first when the cache has nothing for this model (oklm tune)

Sampling

run, chat, serve, describe and agent share these; in serve they are the defaults of what a request leaves out (inference.md, server.md).

Option Default Meaning
-n N 128 (run), 256 (describe); none in serve tokens generated at most. In serve: the length of a request that gives none (max_tokens, max_completion_tokens), which otherwise runs until its sequence is full
--temp T 0.8 temperature; 0 is greedy
--top-k K 40 only the K likeliest tokens are kept
--top-p P 0.95 the likeliest tokens whose probabilities sum to P are kept
--min-p P 0.05 tokens below P times the likeliest one's probability are dropped
--repeat-penalty R 1 penalty on tokens already written; 1 is off
--seed S 1234; in serve, a new one each request, from the clock the sampler's seed. In serve: the seed of every request that names none, the same for each, so that the same request gives the same text

Mixture-of-experts models

run, chat, ppl and serve take these for a MoE model (moe.md); plan takes --n-cpu-moe.

Option Meaning
--expert-cache GB|auto (CPU) the routed experts read from the file into a per-layer LRU cache of that size instead of being held resident; auto: the largest that fits beside the dense part
--pin-gb GB that much of the cache given to the most routed experts of the usage file, never evicted
--expert-usage FILE the routing counts file (default ~/.cache/oklm/usage/<hash>.json)
--expert-direct, --no-expert-direct O_DIRECT reads of the experts, or buffered ones
--expert-loaders N threads that read the missing experts while the resident ones compute (4; 0: none)
--expert-predict last|router the next layer's experts read ahead: its last routing, or its router applied to the current state
--expert-pack FILE each missing expert read in one read from an expert pack (oklm expert-pack)
--expert-mirror FILE every other read sent to a copy of the experts' file on another drive
--expert-stats hit rates and reads printed at exit
--n-cpu-moe N|all, -ncmoe (device backend) the routed experts of the first N layers, or of all, stay in host memory and run on the CPU, the rest of those layers on the device
--expert-gpu GB|auto (device backend) a device tier of that size holding the hottest experts of the --n-cpu-moe layers

--expert-loaders, --expert-predict and --expert-direct, when not given, take what oklm tune --expert-cache kept for this model, cache size and disk.

Speculative decoding

run and serve (inference.md). The answers are those of plain decoding; only the speed changes.

Option Meaning
--draft MODEL, -md a smaller model of the same vocabulary proposes tokens, --draft-n N (5) a step at most
--mtp, --mtp-model FILE (run) the model's own multi-token prediction block as the draft, from the model's file or from another file
--spec-ngram K prompt lookup: up to K tokens that followed the last tokens earlier in the text
--spec-grammar the text a grammar forces (a key, the fixed part of a tool call) proposed as drafts
--spec-guard RATE with --expert-cache: no guessed drafts while the expert cache's recent hit rate is below RATE

Retrieval in generation

run, chat and serve answer from a retrieval index with --rag (rag.md); the agent has its own options, below.

Option Default Meaning
--rag INDEX none an index made by oklm rag index: each question is searched in it and answered from the passages found
--rag-k N 5 passages kept
--rag-mode dense|sparse|hybrid hybrid vectors, keywords (BM25), or both fused
--rag-crag on|off on corrective retrieval: each passage judged, the question rewritten when none is relevant
--rag-route on|off off the model decides between no retrieval, one search and several
--rag-rewrite on|off on (chat, serve) a follow-up rewritten as a standalone query before it is searched
--rag-followup rewrite|join|both both (chat, serve) what a follow-up searches: the rewrite, the previous question joined with it, or both
--rag-rerank MODEL, --rag-rerank-top N none, 20 the candidates reordered by a reranker before the best are kept
--rag-graph on|off off the index's graph walked and fused with the search (an index built with --graph)
--rag-critique on|flag|off off the answer checked against its sources: on generates it once more when a cited statement is unsupported, flag only flags it
--rag-log FILE none every retrieval appended to FILE as a JSON line

Output for programs

Option Meaning
--json the result as one JSON object instead of text, where a command lists it
--progress-json FILE|- the run as JSON lines, one event a line, in FILE; - is standard output, the human report then going to standard error (convert, quantize, ppl, train, rl, abliterate, refusals, rag index; training.md, convert.md)
--dry-run (convert, quantize) say the size the file would have, and write nothing

Other names

Options spelled as other programs spell them are read as OKLM's:

Other name Read as
--model -m
-ngl --gpu-layers
-np --parallel
-md --draft
-ncmoe --n-cpu-moe
--cache-type-k, --cache-type-v -ctk, -ctv
--system-prompt, --permission-mode (agent) --system, --mode
--allowed-tools, --disallowed-tools (agent) --allowedTools, --disallowedTools

The program

oklm version
oklm devices
oklm types
oklm help

version prints the library's version. devices lists the compute backends built in and found, the CPU's features and the thread count. types lists the tensor element types and what OKLM can do with each (read, compute, write). help prints the usage text. None takes an option. See build.md and core.md.

Files

oklm info

oklm info [-t] [--json] <model.gguf>

Prints a model file's metadata and the compatibility verdict: whether OKLM runs it, and why not when it does not (convert.md).

Option Meaning
-t every tensor too: name, type, shape
--json one object, with the layout by block

oklm convert

oklm convert <hf_dir> --out <model.gguf> [--type f32|f16|bf16|q8_0|q4_0|q4_1|q5_0|q5_1] [--name NAME] [-v] [--dry-run]

Converts a Hugging Face checkpoint (config.json and safetensors) to a GGUF file (convert.md).

Option Meaning
--out FILE the GGUF file written (required)
--type T the matrices' type: f32, f16 (default), bf16, q8_0, q4_0, q4_1, q5_0, q5_1
--name NAME general.name (default: the directory's name)
-v each tensor as it is written
--dry-run, --progress-json FILE|- Output for programs

oklm quantize

oklm quantize [--pure] [--imatrix FILE] [--override-kv KEY=TYPE:VALUE] [--dry-run] <in.gguf> <out.gguf> <type>

Requantizes a GGUF model to a llama-quantize type (Q4_K_M, Q8_0, IQ4_XS...; the usage error lists them all) (convert.md).

Option Meaning
--pure every matrix in the base type, none kept at a larger one
--imatrix FILE an importance matrix (oklm imatrix) that weighs the rounding error
--override-kv KEY=TYPE:VALUE a metadata key set in the result (TYPE str, int, float or bool; repeatable)
--dry-run, --progress-json FILE|- Output for programs

oklm imatrix

oklm imatrix -m <model> -f <text.txt> -o <out.imatrix.gguf> [-c 512] [--chunks N] [--process-output]

Computes an importance matrix for oklm quantize --imatrix from a calibration text, in llama-imatrix's format (convert.md).

Option Meaning
-f FILE the calibration text
-o FILE the importance matrix written
-c N the length of each chunk in tokens (512)
--chunks N chunks read at most (default: the whole text)
--process-output the output matrix collected too

oklm split, oklm merge

oklm split <in.gguf> <out-prefix> --split-max-size N[K|M|G] | --split-max-tensors N
oklm merge <shard.gguf> <out.gguf>

split writes a model as shards <out-prefix>-0000i-of-0000N.gguf; merge joins the shards of a split model, given any one of them, into one file (convert.md).

Option Meaning
--split-max-size N[K|M|G] a shard's largest size
--split-max-tensors N a shard's largest tensor count

oklm lora-merge

oklm lora-merge <model.gguf> <adapter.gguf> <out.gguf> [--scale 1] [--save-type bf16]

Merges a LoRA adapter (one oklm train --lora wrote, or llama.cpp's format) into the model: W + scale x alpha / rank x B A (training.md).

Option Meaning
--scale S the adapter's weight (1)
--save-type bf16 the merged weights and the other F32 / F16 matrices written in BF16

oklm expert-pack

oklm expert-pack -m <model.gguf> [-o FILE]

Writes an expert pack of a MoE model: each routed expert's three matrices in one aligned record, which --expert-pack FILE reads in one read (moe.md). -o FILE names the pack (default <model.gguf>.experts).

oklm tokenize

oklm tokenize --model <model.gguf|tokenizer.json|dir> [--bos] [--no-special] [--pieces] <text>
oklm tokenize --model <...> --decode <id,id,...>

Prints the token ids of a text, or the text of token ids, with the tokenizer of a model file, a tokenizer.json or a checkpoint directory (tokenizer.md).

Option Meaning
--bos the tokenizer's special tokens added (BOS, EOS) as the model expects them
--no-special special tokens written in the text are not recognized, they are tokenized as text
--pieces each id with its text
--decode id,id,... the other way: the text of those ids

oklm train-tokenizer

oklm train-tokenizer --vocab <N> --out <tok.gguf> [--min-freq F] [--pre gpt-2|llama-bpe|qwen2|smollm]
    [--special TEXT]... [--chat-template FILE] [-v] <files...>

Trains a byte-level BPE tokenizer on text files and writes it as a vocabulary-only GGUF (tokenizer.md).

Option Meaning
--vocab N the vocabulary's size, special tokens included
--out FILE the file written
--min-freq F the merges stop when the most frequent pair occurs fewer than F times (2)
--pre NAME the pre-tokenizer's pattern set: gpt-2 (default), llama-bpe, qwen2, smollm
--special TEXT a special token (repeatable, 16 at most)
--chat-template FILE a Jinja chat template kept in the file
-v a log line every 500 merges

oklm ingest

oklm ingest <document>...

Prints documents as the text a model reads: text, Markdown, CSV, JSON indented, HTML rendered, DOCX, PDF (multimodal.md). It takes no option; oklm run --doc FILE puts the same text before a prompt.

Models and memory

oklm models

oklm models [list] [--category chat|code|vision|audio|speech|security|retrieval] [--installed] [--json]
oklm models pull NAME... [--accept-licence]
oklm models rm NAME... | models path NAME [--mmproj] | models dir

The catalogue of models OKLM offers to download (it ships none): list shows them, pull downloads, rm removes, path prints an installed model's file and dir the model directory (models.md).

Option Meaning
--category C (list) one category: chat, code, vision, audio, speech, security, retrieval
--installed (list) the installed models only
--json (list) the list as JSON
--accept-licence (pull) answers yes to a licence that has to be agreed to
--mmproj (path) the model's projector file instead of the model's

oklm plan

oklm plan -m <model.gguf> [-c CTX] [--parallel N] [-ctk TYPE] [-ctv TYPE] [--gpu-layers N] [--n-cpu-moe N|all] [--backend cpu|opencl|vulkan|auto]
    [--draft MODEL] [--mmproj FILE] [--gpu-stream N] [--json]   (aliases: --model, -np, -ngl, -ncmoe, -md)
    [--serve [--kv-mem MB] [--tiles SIDE] [--transcribe-model DIR [--transcribe-backend cpu|opencl|vulkan|auto]]
    [--speak-model FILE [--speak-vocoder FILE] [--speak-backend cpu|opencl|vulkan|auto]]
    [--embed-model FILE [--embed-backend B]] [--rerank-model FILE [--rerank-backend B]]]
oklm plan --list [DIR] [--recommend] [-c CTX] [--backend cpu|opencl] [--n-cpu-moe N|all] [--json]

Says the memory a model would take, from its headers alone, against the available RAM and the device's free memory: fits, fits with N layers on the device, or does not fit (exit status 0 when it fits, 3 when it does not). plan --list does it for every GGUF model under models/ or DIR (memory-plan.md).

Option Meaning
-c, -ctk, -ctv, --backend, --gpu-layers, --gpu-stream, --n-cpu-moe what the run would be given (Where it computes)
--parallel N, -np the sequences a server would open
--serve project the model as oklm serve runs it: one cache shared by the requests, what is left beside the model on a device, four contexts on the CPU; the verdict then also says with a smaller cache (one context at least, less than four), and does not fit when the cache holds less than one context (memory-plan.md)
--kv-mem MB with --serve: the size of that cache, as oklm serve --kv-mem takes it (-1: the whole context for each sequence)
--draft MODEL, --mmproj FILE a draft model and a projector counted beside the model
--tiles SIDE, --transcribe-model DIR, --transcribe-backend B, --speak-model FILE, --speak-vocoder FILE, --speak-backend B, --embed-model FILE, --embed-backend B, --rerank-model FILE, --rerank-backend B with --serve: what the server loads beside the model, as oklm serve takes these options, each part counted where the server places it and taken out of the cache's room first (memory-plan.md)
--list [DIR] every model under models/ or DIR, with its need at 8192 tokens and its verdict
--recommend (--list) names the largest model that fits resident and the largest that needs the expert cache
--json one object, with each layer's placement (memory-plan.md)

oklm tune

oklm tune -m <model.gguf|hf_dir> [--backend cpu|opencl|vulkan|auto] [--budget S] [-n TOKENS] [--force] [--cache-dir DIR]
    [--expert-cache GB|auto | --streaming] [--gen TOKENS]

Tunes the CPU kernels' work split for a model on this machine and keeps the result in a cache that contexts apply (OKLM_TUNE=0 turns that off). With --expert-cache it tunes a MoE model's streaming settings instead (moe.md, kernel-autotune-plan.md).

Option Meaning
--budget S seconds the search may take (60); the best found so far is kept
-n TOKENS prompt tokens timed per run (512; 256 for the streaming settings)
--gen TOKENS (--expert-cache) the tokens decoded in each timed run (64)
--force tune again although the cache has a result
--cache-dir DIR where the results are kept (default $XDG_CACHE_HOME/oklm/tuning, else ~/.cache/oklm/tuning)
--expert-cache GB|auto, --streaming time --expert-loaders, --expert-predict and --expert-direct with that cache (--streaming: auto) and keep the fastest with the same output

Inference

oklm run

oklm run -m <model.gguf|hf_dir> -p <prompt> [--doc FILE] [-n 128] [--temp 0.8] [--top-k 40] [--top-p 0.95] [--min-p 0.05] [--repeat-penalty 1]
    [--seed S] [-c CTX] [--backend cpu|opencl|vulkan|auto] [--gpu-layers N] [--tune]
    [--grammar FILE.gbnf | --json-schema FILE.json | --json] [--draft MODEL | --mtp | --mtp-model FILE | --spec-ngram K]
    [--draft-n 5] [--spec-grammar] [--spec-guard RATE] [--observe FILE [--observe-attn all|L,L...] [--observe-lens K]]
    [--ablate L:attn|L:ffn|L:head:H[,...] | --ablate-scan blocks|heads:L] [--ablate-mean]
    [--lora ADAPTER [--lora-scale 1]] (also chat, ppl, serve: a LoRA adapter merged in as the model loads)

Completes a prompt and prints the text (inference.md). It takes the model, placement, sampling, MoE, speculative decoding and retrieval options, and:

Option Meaning
-p PROMPT the prompt
--doc FILE a document whose text goes before the prompt, read as oklm ingest reads it
--grammar FILE.gbnf the output constrained by a GBNF grammar (inference.md)
--json-schema FILE.json the output constrained to the values of a JSON schema
--json the output constrained to any JSON value
--observe FILE a JSON object a line for the prompt's last token then each generated one: what each block wrote, the experts picked (observe.md)
--observe-attn all|L,L... with it, the attention of the last token over the positions, for every block or those listed
--observe-lens K with it, the K likeliest tokens the output head gives after each block
--ablate L:attn|L:ffn|L:head:H[,...] those parts taken out of the run, and what that does to the next token said (observe.md)
--ablate-scan blocks|heads:L every block's two parts, or every head of block L, taken out in turn, a line each
--ablate-mean a part taken out is replaced by its mean over the prompt, not by zero

oklm chat

oklm chat (-m <model.gguf|hf_dir> | --url URL [--api-key KEY] [--remote-model NAME]) [--system TEXT] [--chat-template FILE|NAME] [--template-vars JSON]
    [--tools FILE] [--tool-choice auto|required]
    [--tool-format hermes|llama3|mistral|glm45|glm4|harmony|lfm2|gemma4] [--temp T] [--top-k K] [--top-p P] [--min-p P] [--repeat-penalty R]
    [-c CTX] [--backend cpu|opencl|vulkan|auto] [--gpu-layers N] [--resume] [--kv-save] [--chat-file FILE] [--stats] [--tui]
    [--mmproj FILE|NAME|none [--tiles SIDE] [--fps 1] [--frames 8] [--ffmpeg]]
    [--auto-compact on|off] [--auto-continue on|off] [--continue-max 8]

An interactive chat on the terminal, with a model in this process (-m) or a server's (--url). A line starting with / is a command (/help lists them: /system, /reset, /model, /context, /attach, /save, /export, /tui, /exit...; server.md has the table shared with the window and the dashboard). It takes the model, placement, sampling, MoE and retrieval options, and (inference.md):

Option Meaning
--url URL chat with the model of an OpenAI-compatible server (oklm serve, llama.cpp's server, another) in place of one in this process: http://host:port or https://host:port, with or without /v1; any other path is the API's base and stands for /v1 and an endpoint given whole (.../v1/chat/completions) is read as its base; OpenAI, Gemini, Anthropic, Mistral, DeepSeek and Ollama are reached that way (gui.md). Not with -m. Without -m and --url, the baseUrl setting is the server. What each command and option does there, and which need -m: inference.md
--api-key KEY with --url: the key sent (Authorization: Bearer); else OKLM_API_KEY, else OPENAI_API_KEY
--remote-model NAME with --url: the server's model. Without it, the one the server lists (GET /v1/models); among several, a list to choose from on a terminal, and status 2 with the list when the input is not typed
--system TEXT the system prompt
--tools FILE a JSON file of tool definitions the model may call; /tool RESULT answers a call (inference.md)
--tool-choice auto|required whether the model may answer in text or must call a tool
--tool-format NAME the wire format of tool calls, in place of the template's: hermes, llama3, mistral, glm45, glm4, harmony, lfm2, gemma4
--resume go on with the conversation kept in .oklm/chat/last.json
--chat-file FILE keep the conversation in that file instead
--kv-save keep the cache beside the conversation (.kv), so --resume starts warm (inference.md)
--auto-compact on|off near the context's end the older messages are summarized into a note by themselves (on), or only with /compact; for this run, over autoCompact of the user's settings, which /autocompact writes and oklm agent shares (inference.md)
--auto-continue on|off a reply the context or the token limit cut is gone on with by itself (on), or only with /continue; for this run, over autoContinue
--continue-max N the pieces a reply is continued with in a row at most (8); for this run, over continueMax
--stats each turn's cached and evaluated tokens
--tui the full-screen UI: Markdown, folded reasoning, a status bar
--mmproj FILE|NAME|none the projector: /attach FILE then takes a picture, a sound or a video; none loads none (multimodal.md)
--tiles SIDE, --fps F, --frames N, --ffmpeg how attached pictures and videos are read, as oklm describe
--observe FILE as oklm run

Without --resume, --kv-save or --chat-file nothing is written.

oklm ppl

oklm ppl -m <model.gguf> -f <text.txt> [-c CTX] [--progress-json FILE|-]

Perplexity of a model over a text file, the second half of each window scored (inference.md). It takes placement, MoE and --lora options, -f FILE (the text) and --progress-json.

oklm embed

oklm embed -m <model.gguf|hf_dir> [--json] [--query] [--pooling cls|mean|last] [--no-normalize] [--query-prefix TEXT]
    [--passage-prefix TEXT] [-c CTX] [-b BATCH] [--backend cpu|opencl|vulkan|auto] [--gpu-layers N] <text>... | --file F

One embedding a text, or a line of --file F, from a text encoder or a decoder embedding model: one vector a line (rag.md).

Option Meaning
--json OpenAI's /v1/embeddings response
--query the texts are queries: the query prefix is put before them, not the passage prefix
--pooling cls|mean|last how the tokens' vectors become one, in place of what the model says
--no-normalize the vectors not L2-normalized
--query-prefix TEXT, --passage-prefix TEXT the text put before a query or a passage, in place of the model's
-b BATCH texts evaluated at once
--file F the texts, one a line

oklm rerank

oklm rerank -m <model.gguf|hf_dir> -q QUERY [--top N] [--json] [-c CTX] [--backend cpu|opencl|vulkan|auto] <passage>... | --file F

Scores passages against a query and prints them best first: a cross-encoder gives its relevance logit, a chat model answers whether the passage is relevant (rag.md).

Option Meaning
-q QUERY, --query QUERY the query
--top N the N best only
--json /v1/rerank's response
--file F the passages, one a line

Training

oklm train

oklm train (-f <text.txt> | --chat-data <conv.jsonl>) --out <prefix> [-m <model> [--lora R ...] | --resume <ckpt.gguf> | --tokenizer <tok> | --vocab N]
    [--arch gpt2|llama|qwen2|qwen3] [--layers 4] [--embd 128] [--heads 4] [--kv-heads H] [--head-dim D] [--ff N]
    [--rope-theta 10000] [--ctx 256]
    [--steps 1000] [--batch 4] [--accum 1] [--seq T] [--lr 3e-4] [--lr-min 3e-5] [--warmup 50] [--wd 0.1] [--clip 1] [--dropout 0] [--seed S]
    [--val-frac 0.1] [--eval-every 100] [--eval-batches 8] [--save-every N] [--sample-every N] [--sample-len 200]
    [--backend auto|cpu|opencl|rocm] [--recompute]
    [--quantized-base] [--act-type f32|bf16] [--opt-type f32|bf16] [--save-type f32|bf16] [--matmul-precision f32|bf16|f16]
    [--grad-type f32|bf16] [--loss-scale dynamic|N|off]

Trains a GPT-2, Llama or Qwen model from scratch, fine-tunes a model (-m), trains LoRA adapters on it (--lora), or continues a checkpoint (--resume). Checkpoints are <prefix>-<step>.gguf with a .train sidecar of the optimizer's state (training.md).

Option Default Meaning
-f FILE the training text
--chat-data FILE conversations as JSON lines instead, the loss on the assistant turns (training.md)
--out PREFIX the checkpoints' prefix; its directories are made when missing
-m MODEL the model fine-tuned
--resume FILE the checkpoint (or with -m, the adapter) continued; its options come from the sidecar
--tokenizer FILE, --vocab N from scratch: a tokenizer file, or a BPE tokenizer of N tokens trained on the text
--arch gpt2|llama|qwen2|qwen3 gpt2 from scratch: the architecture
--layers, --embd, --heads, --ctx 4, 128, 4, 256 blocks, width, attention heads, trained context
--kv-heads H, --head-dim D, --ff N, --rope-theta T Llama and Qwen: key / value heads, a head's width, the feed-forward width, RoPE's base (10000)
--steps N 1000 updates
--batch N, --seq T, --accum N 4, , 1 rows a batch, tokens a row, batches summed per update
--lr, --lr-min, --warmup 3e-4, 3e-5, 50 the learning rate, its floor at the end of the cosine decay, the warm-up steps
--wd, --clip, --dropout 0.1, 1, 0 weight decay, the gradient's norm limit, residual dropout
--seed S the seed of the initial weights and of the batches
--val-frac F, --eval-every N, --eval-batches N 0.1, 100, 8 the text's last share kept for validation, how often and over how many windows its loss is measured
--save-every N a checkpoint every N steps (and one at the end)
--sample-every N, --sample-len L off, 200 every N steps and at the end, a text of L tokens the model writes from its end-of-text token
--lora R, --lora-alpha A, --lora-targets LIST rank-R adapters on -m's projections (q,k,v,o,gate,up,down, or all), the model frozen; the checkpoints are adapter files
--quantized-base with --lora: the frozen matrices kept in -m's types instead of F32
--recompute each block's activations recomputed in the backward pass: far less activation memory, about one more forward pass
--act-type, --opt-type, --grad-type f32|bf16 f32 the stored activations, AdamW's moments, the gradients held between passes, in BF16 to halve their memory (training.md)
--save-type f32|bf16 f32 the checkpoints' matrices
--matmul-precision f32|bf16|f16 f32 (--backend rocm) the inputs of the large products rounded to that type
--loss-scale dynamic|N|off off the gradients computed on the loss times a scale and divided back, for F16 products
--backend, --progress-json Where it computes, Output for programs

oklm rl

oklm rl (-m <model> | --resume <ckpt.gguf>) -f <tasks.jsonl> --out <prefix> [--steps 100] [--tasks 8] [--group 8]
    [--max-new 32] [--temp 1] [--lr 1e-5] [--wd 0] [--val-frac 0.1] [--eval-every 20] [--save-every N] [--seed S]
    [--updates 1] [--clip 0.2] [--kl 0] [--ref <model>] [--backend auto|cpu|opencl|rocm]

Reinforcement-learning fine-tuning on tasks with known answers (tasks.jsonl: one {"prompt": ..., "answer": ...} a line). A completion earns 1 when its last number is a numeric answer, or when it starts with any other answer; the advantages are relative to the group (training.md).

Option Default Meaning
-m MODEL, --resume FILE the model, or a checkpoint continued
-f FILE, --out PREFIX the tasks, the checkpoints' prefix
--steps N, --tasks N, --group N 100, 8, 8 updates, tasks a step, completions sampled per task
--max-new N, --temp T 32, 1 a completion's length and sampling temperature
--lr, --wd, --seed 1e-5, 0 as oklm train
--val-frac F, --eval-every N, --save-every N 0.1, 20 the tasks held out, how often they are scored, how often a checkpoint is written
--updates N, --clip C 1, 0.2 updates from each sampled batch, the ratio clipped to 1 +/- C past the first
--kl W, --ref MODEL 0 the weight of a KL penalty against a reference model (default: the starting model)

oklm abliterate

oklm abliterate -m <model> --harmful <refused.txt> --harmless <answered.txt> [--out <out.gguf>] [--layer auto|N]
    [--band 0.3,0.7] [--scale 1] [-n PROMPTS] [--system TEXT] [--template-vars JSON] [--save-dirs FILE]
oklm abliterate -m <model> --load-dirs FILE --layer N --out <out.gguf> [--scale 1]

Finds a chat model's refusal direction per layer from two prompt sets and, with --out, projects it out of the weights (training.md).

Option Default Meaning
--harmful FILE, --harmless FILE prompts the model declines, prompts it answers, one a line
--out FILE the model written; without it the directions are only measured
--layer auto|N auto the layer whose direction is removed; auto: the best within --band
--band LOW,HIGH 0.3,0.7 the share of the depth auto looks in
--scale S 1 how much of the direction is removed (0 to 2)
-n PROMPTS all the first PROMPTS lines of each file only
--system TEXT, --template-vars JSON the system prompt and the template variables the prompts are rendered with
--save-dirs FILE, --load-dirs FILE keep the directions; write another layer or scale from them without recomputing

oklm refusals

oklm refusals -m <model> [--prompts <refused.txt>] [--ref <model> --harmless <answered.txt>] [-n 48] [--limit N]
    [--backend cpu|opencl|vulkan|auto] [--system TEXT] [--template-vars JSON] [--show]

Answers each prompt of a file greedily and counts the answers that read as refusals; with a reference model, the first-token KL divergence from it on prompts the model should answer (training.md).

Option Default Meaning
--prompts FILE the prompts judged
--ref MODEL --harmless FILE the reference model and the prompts the divergence is measured on
-n N 48 tokens of each answer
--limit N all the first N prompts only
--show each answer printed as it is judged (REFUSED or answered)
--system TEXT, --template-vars JSON as oklm abliterate

Serving

oklm serve

oklm serve -m <model> [--host 127.0.0.1] [--port 7430] [--tls-cert FILE --tls-key FILE] [--api-keys FILE] [--insecure] [--fixed-context] [--fixed-model]
    [--parallel N] [--queue N] [--workers N] [--kv-mem MB] [--prefill-chunk N] [--prefill-share F] [--deterministic] [--observe] [--observe-lens K] [--spec-ngram K] [--model-id NAME] [-c 4096] [--backend cpu|opencl|vulkan|auto] [--temp T] [--top-k K] [--top-p P]
    [--min-p P] [--repeat-penalty R] [--seed S] [-n N]
    [--mmproj FILE [--tiles SIDE] [--ffmpeg]] [--draft MODEL [--draft-n 5]] [--swap-mem MB] [--lora FILE [--lora-scale S]] [--chat-template FILE|NAME]
    [--lora-module NAME=ADAPTER.gguf ...] [--spec-grammar] [--spec-guard RATE] [--prefix-dir DIR [--prefix-min N] [--prefix-disk-mb MB]]
    [--embed-model EMB [--embed-model-id NAME] [--embed-backend cpu|opencl|vulkan|auto]]
    [--rerank-model R [--rerank-model-id NAME] [--rerank-backend cpu|opencl|vulkan|auto]]
    [--transcribe-model W [--transcribe-model-id NAME] [--transcribe-backend BACKEND]] [--upload-max MB] [--ingest-keep SECONDS]
    [--speak-model S [--speak-vocoder FILE] [--speak-voice FILE] [--speak-voices DIRS] [--speak-model-id NAME] [--speak-backend BACKEND] [--speak-max BYTES]]
    [--preload]

An HTTP server: OpenAI-compatible and Anthropic-compatible routes, a dashboard on /, requests sharing one context in batches. Beyond loopback, TLS and API keys are required. server.md has every option with its default (Running it) and every route (Endpoints). Besides the model, placement, sampling, MoE, speculative decoding and retrieval options:

Option Meaning
--host ADDR, --port N where it listens (127.0.0.1, 7430)
--tls-cert FILE --tls-key FILE HTTPS
--api-keys FILE the accepted keys, one a line
--insecure a non-loopback address allowed without TLS or keys
--fixed-context, --fixed-model POST /api/context and POST /api/model refused
--parallel N, --queue N, --workers N requests generating at once, requests waiting, connections served
--kv-mem MB the paged KV cache's size (0: what fits; -1: the contiguous cache)
--swap-mem MB host memory for the rows of preempted requests
--prefill-chunk N, --prefill-share F prompt tokens of one request evaluated in a step, and their share of a step while others decode
--deterministic the same tokens at temperature 0 whether the prompt was cached or not (server.md)
--prefix-dir DIR, --prefix-min N, --prefix-disk-mb MB the prefix cache kept on disk between runs (server.md)
--observe, --observe-lens K activations recorded, read with GET /api/activations (server.md)
--model-id NAME the model's name in the API
--lora-module NAME=ADAPTER.gguf an adapter served beside the model, used by the requests whose model is NAME (repeatable)
--mmproj FILE, --tiles SIDE, --ffmpeg chat messages take images, videos and sounds (server.md)
--embed-model EMB, --embed-model-id NAME, --embed-backend B /v1/embeddings (server.md)
--rerank-model R, --rerank-model-id NAME, --rerank-backend B /v1/rerank (server.md)
--transcribe-model W, --transcribe-model-id NAME, --transcribe-backend B /v1/audio/transcriptions and /v1/audio/translations (server.md)
--speak-model S, --speak-vocoder FILE, --speak-voice FILE, --speak-voices DIRS, --speak-model-id NAME, --speak-backend B, --speak-max BYTES /v1/audio/speech and /v1/audio/voices (server.md)
--preload the transcription and speech models loaded at start, not by the first request that needs them
--upload-max MB, --ingest-keep SECONDS the largest document or sound sent as a body (32), how long a conversation's documents are kept (3600) (server.md)
--rag-dir DIR a directory of indexes a request may name (server.md)

oklm serve --list, --stop

oklm serve --list [--json] | --stop PORT|PID

The servers of this user that run (each registers itself at start: pid, port, model, command), --json as JSON; --stop ends one by its port or its pid (server.md).

oklm serve --keep

oklm serve --keep NAME -m <model> [the options of serve] | --rekeep NAME ... | --kept [--json] | --unkeep NAME

A server kept across logins, as a service of the user (server.md).

Option Meaning
--keep NAME ... the command that follows kept under NAME and started now, at each login and again when it fails
--rekeep NAME ... the command kept under NAME changed
--kept the kept servers with their state (--json: as JSON)
--unkeep NAME one stopped and its unit deleted
--start-kept NAME a kept server that does not run started, as its service starts it (what Windows runs at login)

Agent

oklm agent

oklm agent (-m <model> [--chat-template FILE|NAME] | --url URL [--api-key KEY] [--remote-model NAME] | --mock FIXTURE.jsonl) [-p TASK] [--project DIR]
    [--mode default|acceptEdits|plan|bypassPermissions] [--dangerously-skip-permissions] [--allow RULE] [--ask RULE] [--deny RULE]
    [--max-turns 32] [--max-tokens 8192] [--max-background 2] [--no-sandbox] [--allow-network] [--keep-services] [--limit NAME=VALUE]
    [--auto-compact on|off] [--auto-continue on|off] [--continue-max 8]
    [--thinking hide|show] [--system TEXT] [--append-system-prompt TEXT] [--output-style default|explanatory|learning|NAME]
    [--no-memory] [--save-training FILE] [--no-tool-grammar] [--debug] [--worktree [NAME]] [--temp T] [-c 32768]
    [--backend cpu|opencl|vulkan|auto] [--tool-rag K] [--rag INDEX | --rag-project] [--rag-memory] [--rag-inject off|always|route]
    [--resume [SESSION]] [--fork-session] [--kv-save]
    [--output-format text|json|stream-json] [--input-format text|stream-json] [--listen [PORT]] [--allowedTools RULES]
    [--disallowedTools RULES] [--add-dir DIR] [--mcp-config FILE [--strict-mcp-config]]

An agent that reads, searches and edits a project's files and runs commands within permission rules: a REPL on the terminal, or with -p one task and an exit. user-guide.md is the guide, agent-loop.md the loop and the headless use, commands.md its slash commands. It takes the model, --backend, -c (32768) and --temp, and:

Option Meaning
--url URL, --api-key KEY, --remote-model NAME the model of an OpenAI-compatible server instead of one in process (agent-loop.md)
--mock FIXTURE.jsonl a scripted model, for tests
-p TASK one task, without a terminal; the exit status is 0 when it succeeded, 1 when it ended in an error or was not answered for want of room in the context (agent-sessions.md)
--project DIR the project's root (default: the current directory)
--mode MODE the permission mode: default, acceptEdits, plan, bypassPermissions (agent-permissions.md)
--dangerously-skip-permissions the bypass mode
--allow RULE, --ask RULE, --deny RULE a permission rule for the session (repeatable)
--allowedTools RULES, --disallowedTools RULES several allow or deny rules in one argument, commas or blanks between them
--add-dir DIR another working directory
--max-turns N, --max-tokens N model turns a request may take (32), tokens of a turn (8192). Without --max-tokens the 8192 are the size of one piece of an answer, which goes on past them; given, they end an answer for good
--auto-compact on|off, --auto-continue on|off, --continue-max N compaction near the window's end, and going on with an answer the token limit or the context cut, by themselves (on, the default) or only with /compact and /continue; the pieces in a row at most (8). Over the settings autoCompact, autoContinue, continueMax (agent-sessions.md)
--max-background N background sub-agents at once (2) (agent-tools.md)
--no-sandbox, --allow-network commands not confined; the network kept in the sandbox (agent-permissions.md)
--limit NAME=VALUE a resource limit of the commands: cpu=S, memory=8G, file=4G, files=N, procs=N
--keep-services the service jobs left running when the session ends (processes.md)
--thinking hide|show the model's reasoning folded or shown live
--system TEXT, --append-system-prompt TEXT the system prompt replaced, or text added to it
--output-style NAME default, explanatory, learning, or a style of the project
--no-memory the OKLM.md memory files left out (agent-settings.md)
--save-training FILE the run appended as a JSONL training example
--no-tool-grammar tool calls not constrained by the tools' grammar
--debug the debug log written (agent-settings.md)
--worktree [NAME] work in a git worktree of its own (git.md)
--resume [SESSION], --fork-session resume the latest session, or the one named; as a copy in a new session (agent-sessions.md)
--kv-save the in-process model's cache kept beside the transcript
--output-format text|json|stream-json, --input-format text|stream-json the headless formats (agent-loop.md)
--listen [PORT] serve the agent over HTTP on loopback for other programs (agent-loop.md)
--mcp-config FILE, --strict-mcp-config MCP servers from that file; only those (agent-extensions.md)
--tool-rag K the K tools, skills and sub-agent types that fit each message retrieved, the others deferred (rag.md)
--rag INDEX, --rag-project, --rag-memory the Retrieve tool over an index, over the project's documentation, over the earlier sessions too (rag.md)
--rag-inject off|always|route passages given with each message, always or when the router says so
--rag-critique on|flag|off the final answers checked against their sources

oklm agent config, skills, sessions

oklm agent config get KEY | set KEY VALUE | unset KEY [--scope local|project|user] [--project DIR]
oklm agent skills validate PATH... | agent skills list [--project DIR]
oklm agent sessions [list] [--json] | export ID [--format markdown|json] [-o FILE] | rm ID [--project DIR]

config reads a setting (the value in force) or writes it in one settings file (agent-settings.md). skills validate checks skills as the agent loads them, skills list lists those it finds (agent-tools.md). sessions lists the project's sessions, writes one as Markdown or JSON, or removes one (agent-sessions.md).

Option Meaning
--scope local|project|user (config) the settings file: the project's personal one (default), the project's, the user's
--project DIR the project's root
--json (sessions list) the list as JSON
--format markdown|json, -o FILE (sessions export) the format, the file written

oklm agent plugin, mcp, commands, agents

oklm agent plugin ... | mcp ... | commands [--markdown] | agents | attach ID | send ID MESSAGE | stop ID
oklm agent --stop-listening PORT

Each of these prints its own usage when given nothing.

Command What it does
agent plugin list | install SOURCE [--scope user|project] | remove | enable | disable NAME | marketplace add NAME SOURCE | marketplace list plugins and marketplaces (agent-extensions.md)
agent mcp list | get NAME | add NAME [--scope S] [--transport stdio|http|sse] [--env K=V] [--header 'K: V'] -- COMMAND | URL | add-json NAME JSON | remove NAME | login NAME | logout NAME MCP servers (agent-extensions.md)
agent commands [--markdown] the slash-command reference; --markdown is commands.md
agent agents | attach ID | send ID MESSAGE | stop ID the background sessions: listed, followed, sent a message, ended (agent-tools.md)
agent --stop-listening PORT asks the agent listening on that port to end: it closes its listener and keeps its session

oklm permissions, oklm audit

oklm permissions [--project DIR] [--mode MODE] <Tool> <argument>
oklm audit [--tool T] [--decision D] [--session S] [log.jsonl ...]

permissions says how the project's rules decide a tool call, without running it: exit status 0 allow, 3 ask, 4 deny (agent-permissions.md). audit lists the permission decisions recorded in .oklm/audit (agent-permissions.md).

Option Meaning
--project DIR, --mode MODE (permissions) the project and the permission mode the call is judged in
--tool T, --decision allow|ask|deny, --session S (audit) only the decisions of that tool, of that kind, of that session

Retrieval

oklm rag index PATH... --out INDEX [--model <model.gguf|hf_dir>|hash[:DIM]] [--query-prefix TEXT] [--passage-prefix TEXT]
    [--backend cpu|opencl|vulkan|auto] [--chunker auto|fixed|recursive|code|semantic] [--chunk-size N] [--overlap N]
    [--vec f32|f16|q8_0] [--no-bm25] [--rebuild] [--graph --gen-model MODEL] [--caption-model MODEL [--caption-mmproj F]]
    [--caption-prompt TEXT|@FILE] [--gen-backend cpu|opencl|vulkan|auto] [--progress-json FILE|-]
oklm rag search INDEX QUERY [-k 10] [--mode dense|sparse|hybrid] [--fusion rrf|weighted] [--rrf-k 60] [--weights D,S]
    [--rerank MODEL [--rerank-top 50]] [--rag-graph on|off [--graph-hops 1|2]] [--json|--json-full]
oklm rag info INDEX [--sources] [--json]
oklm rag eval DIR --index INDEX [--model MODEL] [--query-prefix TEXT] [--mode dense,sparse,hybrid] [--fusion rrf,weighted]
    [--chunk-size N|0] [--depth 100] [--rerank MODEL [--rerank-top 50]] [--run PREFIX] [--json]

A retrieval index of documents in one GGUF file: chunks, their vectors and BM25 postings (rag.md). index builds it or updates it in place (unchanged files kept by their hash), search queries it, info describes it, eval scores retrieval on a BEIR-format set.

Option Default Meaning
--out INDEX (index) the index written or updated
--model MODEL|hash[:DIM] (index, eval) the embedding model, or the hashed bag of words (rag.md)
--query-prefix TEXT, --passage-prefix TEXT the model's the text put before a query or a passage
--chunker auto|fixed|recursive|code|semantic auto how documents are cut
--chunk-size N, --overlap N a chunk's size in tokens, the tokens consecutive fixed windows share
--vec f32|f16|q8_0 f16 the vectors' type
--no-bm25 no keyword postings
--rebuild the existing index ignored
--graph --gen-model MODEL, --gen-backend B the entities and relations a generator extracts from each chunk kept (rag.md)
--caption-model MODEL, --caption-mmproj F, --caption-prompt TEXT|@FILE the figures of documents indexed as captions by a vision model (rag.md)
-k N 10 (search) hits printed
--mode dense|sparse|hybrid (search; eval: a list) vectors, keywords, or both (rag.md)
--fusion rrf|weighted, --rrf-k N, --weights D,S , 60 how the two rankings are fused
--rerank MODEL, --rerank-top N , 50 the best N candidates reordered by a reranker
--rag-graph on|off, --graph-hops 1|2 (search) the index's graph walked, and how far
--sources (info) each source with its hash and state
--index INDEX, --depth N, --run PREFIX , 100 (eval) the index scored, how deep the rankings go, the rankings written in TREC run format
--json, --json-full, --progress-json rag.md

Multimodal and speech

oklm describe

oklm describe -m <model> --mmproj <mmproj.gguf> [-p QUESTION] [--tiles SIDE] [-n 256] [--fps 1] [--frames 8] [--ffmpeg] [--chat-template FILE|NAME]
    [--vision-id] [--then QUESTION [--then-with FILE]...] [--kv-save FILE] <image, sound or video>...
oklm describe -m <model> --mmproj <mmproj.gguf> --kv-load FILE -p QUESTION [<image, sound or video>...] [--then ...] [--kv-save FILE]

Answers a question about images, sounds or videos with a vision or an audio model; the projector file decides which (multimodal.md). It takes sampling, -c, --backend and --chat-template, and:

Option Meaning
--mmproj FILE the projector (required)
-p QUESTION the question (default: "Describe this image.", "this video." or "this sound.")
--tiles SIDE a picture's longest side before it is sliced into tiles (0: the encoder's default)
--fps F, --frames N videos: frames sampled a second (1), frames at most (8) (multimodal.md)
--ffmpeg containers other than Ogg Theora decoded through ffmpeg
--vision-id pictures and videos numbered in the prompt, as Qwen2-VL's template does
--then QUESTION, --then-with FILE a second question after the answer, with media of its own
--kv-save FILE, --kv-load FILE the conversation's cache written to a KV file, and taken back in a later run to ask the next question without evaluating the media again

oklm transcribe

oklm transcribe -m <whisper checkpoint dir> [--language auto|CODE] [--translate] [--timestamps |
    --format text|timed|srt|vtt|json] [--no-timestamps] [--detect-language] <audio file>

Speech to text with a Whisper checkpoint (multimodal.md).

Option Meaning
--language auto|CODE the language spoken; auto (default): found from the first 30 seconds
--translate English written whatever the language
--format text|timed|srt|vtt|json the words (default), a line a segment with its times, subtitles, or JSON
--timestamps --format timed
--no-timestamps decoded without timestamp tokens, a cut every 30 seconds
--detect-language the language and its probability printed, nothing transcribed
--backend B where the model computes (default cpu)

oklm speak

oklm speak -m <speech model> [--vocoder <vocoder.gguf>] [-o speech.wav] [--voice <speaker.json>] [TEXT]

Text to speech into a WAV or FLAC file, with OuteTTS and WavTokenizer's decoder, or Soprano (speech.md). The text is given, or read from standard input.

Option Meaning
--vocoder FILE the model that turns the audio codes into sound (found by itself for a catalogue model)
-o FILE the file written (speech.wav; a name ending in .flac writes FLAC)
--voice FILE a speaker file of OuteTTS in place of the built-in voice
--temp, --top-k, --repeat-penalty, --seed how the audio codes are sampled
--no-guide the first token of each word left to the model
--words N, --prompt-words N, --first-words N words a pass (24), words a prompt (64), words of the first pass (speech.md)
--speed X 0.25 to 4, the pitch kept
--codes FILE audio codes read from a file, the language model skipped
--backend B where the language model computes (default cpu)

oklm-gui

oklm-gui holds the graphical tools in one program; --tool NAME picks one, chat without it (gui.md).

oklm-gui [--tool NAME] [--size WIDTHxHEIGHT] [--screenshot FILE.png]
Option Meaning
--tool NAME the tool; --tool list prints the names, one a line
--size WIDTHxHEIGHT the window's size
--screenshot FILE.png the window saved to the file, then the program ends (for checks without a person)
--help, -h the tool's usage
--tool Window Described in
launcher a button for each tool, and the documentation The launcher
chat a chat with a model the window serves, a server, or a listening agent The chat window
models the catalogue: download, delete, open a tool on a model The model manager
servers start, watch and stop oklm serve The server manager
inspect a model's layers, where each is and the memory it takes The model inspector
convert oklm convert and oklm quantize with the loss measured The conversion tool
train oklm train, rl, refusals and abliterate followed The training window
sessions the agent sessions of a project folder The sessions browser
rag build, check, search and evaluate a retrieval index The retrieval manager

oklm-gui --tool launcher

oklm-gui --tool launcher [--press open:TOOL|docs]...

The home window. It takes no option of its own.

oklm-gui --tool chat

oklm-gui [--tool chat] (-m <model.gguf> [-c 4096] [--backend cpu|opencl|vulkan|auto] [--parallel 2]
    [--gpu-layers N] [--n-cpu-moe N|all] [--expert-gpu GB|auto] [--kv-mem MB] [-ctk TYPE] [-ctv TYPE] [--devices N]
    [--draft MODEL [--draft-n N]] [--spec-ngram N] [--prefix-dir DIR] [--chat-template FILE|NAME] [--tiles SIDE]
    [--expert-cache GB|auto [--pin-gb GB] [--expert-pack FILE]]
    | --url http(s)://host:port [--key KEY] [--ca FILE] [--no-verify] | --agent PORT)
    [--rag INDEX.gguf] [--mmproj FILE] [--attach FILE]... [--ffmpeg] [--embed-model NAME|FILE]
    [--rerank-model NAME|FILE | --no-rerank] [--no-doc-cache]
    [--chat-file FILE.json]
    [--send TEXT]... [--screenshot FILE.png]
    [--auto-compact on|off] [--auto-continue on|off] [--continue-max 8]
    [--transcribe-model NAME|DIR | --no-transcribe] [--transcribe-language CODE] [--transcribe-translate]
    [--transcribe-attached]
    [--listen-silence SECONDS] [--listen-threshold DB] [--listen-send] [--no-partial] [--mic NAME] [--mic-file FILE[@TIME]]...
    [--talk] [--no-interrupt]
    [--speak-model NAME|FILE [--speak-vocoder FILE] [--speak-voice FILE] | --no-speak] [--read-aloud]
    [--speak-speed 0.25..4] [--speak-whole]
    [--speak-file FILE.wav] [--size 1100x720] [--models] [--model-settings] [--settings 'OPTIONS']...
    [--open-servers FILE.png]
    [--press speech|models[:NAME]|connect[:PORT]|link|own|choose[:PORT|:URL|:NAME|:file[:PATH]|:mmproj[:PATH]]|use
    |advanced[:TAB]|continue|context:N|apply-context|paste|fill:FIELD=VALUE|list-models
    |pick-model:NAME|save|forget:URL|model:NAME]...
    [--shot-when resizing|recording|speaking|talking|commands:TEXT]

The chat window (Options has each option in full).

Option Meaning
-m MODEL, -c N, --backend B, --parallel N the model the window serves itself on loopback, its context, backend and sequences
--url URL, --key KEY, --ca FILE, --no-verify a server at an address instead: its API key, a certificate authority file, the certificate not checked
--agent PORT attach to the agent listening on that port (gui.md)
--models open on the Models window
--model-settings, --settings 'OPTIONS' the model settings open at start; options of oklm serve set and applied in them, for checks without a person (gui.md)
--rag INDEX.gguf a retrieval index beside the model (gui.md)
--mmproj FILE, --attach FILE, --ffmpeg the projector, files attached to the first message, videos through ffmpeg (gui.md)
--embed-model M, --rerank-model M, --no-rerank, --no-doc-cache the models that index and order a document too long for the context; its index not kept
--chat-file FILE.json a saved conversation opened
--send TEXT a message sent at start (repeatable)
--transcribe-model M, --no-transcribe, --transcribe-language CODE, --transcribe-translate, --transcribe-attached speech to text (gui.md)
--listen-silence S, --listen-threshold DB, --listen-send, --no-partial, --mic NAME, --mic-file FILE[@TIME] the microphone: the pause that ends a recording, the level that is speech, the words sent at once, no words shown while a recording is still coming, the device, a sound file in its place
--talk, --no-interrupt a spoken conversation; speaking over the voice does not stop it (gui.md)
--speak-model M, --speak-vocoder FILE, --speak-voice FILE, --no-speak, --read-aloud, --speak-speed X, --speak-whole, --speak-file FILE.wav replies read aloud (gui.md)

oklm-gui --tool models

oklm-gui --tool models [--select NAME] [--installed]

--select NAME selects a model of the catalogue at start; --installed lists the installed models only.

oklm-gui --tool servers

oklm-gui --tool servers [-m MODEL] [--backend cpu|opencl|vulkan|auto] [-c 4096] [--profile NAME]
    [--options 'OPTIONS OF oklm serve'] [--save-profile NAME] [--advanced [TAB]]
    [--press start|keep:NAME|profile:NAME|select:T|stop:T|restart:T|unkeep:NAME|rekeep:NAME|
    ask-rekeep:NAME|dashboard:T|chat:T|context:N|yes|no]...

-m, --backend and -c are what the form shows at start; --profile NAME loads a kept profile in it.

Option Meaning
--options 'OPTIONS' a line of options of oklm serve read into the form: each to its control, the rest to the Advanced window's text field (gui.md)
--save-profile NAME the form kept under that name at start
--advanced [TAB] the Advanced window open at start, on the tab named

oklm-gui --tool inspect

oklm-gui --tool inspect [-m MODEL.gguf | --url URL [--key KEY]] [--backend cpu|opencl|vulkan] [-c CTX]
    [--gpu-layers N] [--n-cpu-moe N] [--ram MIB] [--select embeddings|BLOCK]
    [--prompt TEXT [--chat] [--tokens N] [--step N]] [--layer BLOCK] [--view writes|lens|effects]
Option Meaning
-m MODEL.gguf a file, read through its tensor table (no weight is loaded)
--url URL, --key KEY a running server, with its real placement
--backend, -c, --gpu-layers, --n-cpu-moe where the projected placement starts
--ram MIB project against that much host memory
--select embeddings|BLOCK, --layer BLOCK the part shown as clicked; the block whose attention is shown
--prompt TEXT, --chat, --tokens N, --step N a prompt run through the model, as a user's message, the tokens generated (8), the token shown
--view writes|lens|effects what the activation panel shows

oklm-gui --tool convert

oklm-gui --tool convert [-m <model.gguf|hf_dir>] [--out FILE.gguf] [--type Q4_K_M] [--pure] [--imatrix FILE]
    [--name NAME] [--text FILE.txt] [-c 512] [--run] [--check]

The options fill the window's fields (-m, --out, --type, --pure, --imatrix, --name as the commands'; --text FILE.txt and -c for the perplexity measure); --run starts the conversion, --check the measure once the result is there.

oklm-gui --tool train

oklm-gui --tool train [--mode scratch|finetune|lora|chat|rl|advanced] [--set FIELD=VALUE]... [--start]
    [--oklm PROGRAM] [--press stop|merge|export|export:FILE|chat]...

--mode is the tab shown; --set FIELD=VALUE gives a field its value (the usage lists the names); --start starts the form's command at once; --oklm PROGRAM is the oklm the commands run.

oklm-gui --tool sessions

oklm-gui --tool sessions [--project DIR] [--search TEXT] [--session ID]
    [--press select:ID|select:live|attach|resume|export|export:FILE|delete|yes|no]...

--project DIR is the project folder, --search TEXT filters the sessions by a text their messages hold, --session ID opens one.

oklm-gui --tool rag

oklm-gui --tool rag [--index INDEX.gguf] [--tab build|sources|search|evaluate|attach] [--query TEXT]
    [--model NAME|FILE] [--press sources:PATH|build|eval:DIR]...

--index opens an index at start, --tab chooses the tab, --query searches once the index is read, --model chooses the embedding model of the Build tab.