OKLMdocumentation v1.0.3
On this page

OKLM 1.0.0 release notes

What OKLM 1.0.0 holds, what it promises to keep across the 1.x versions, the platforms it was tested on and its known limits.

Status: released on 2026-11-16, tag v1.0.0. The sanitizer and Wine runs that came before the tag are in roadmap.md.

What 1.0.0 is

One source tree in C17 for the whole life of a model, with no external library touching the model: its files, its tokenizer, inference on the CPU and on a GPU, training, serving, an agent, retrieval, images, sounds and speech. Since 0.1.0 (2026-11-02), which released the models trained from scratch, the program itself grew from an engine with an OpenCL backend to what follows. Each line names the page that describes it and holds its measures.

  • Model files. GGUF read and written byte for byte, safetensors read, every ggml quantization type (the legacy ones, the K and I quants, the ternary and 4-bit float types), oklm convert, oklm quantize, oklm imatrix, oklm info, oklm plan (convert.md, formats/quant.md, memory-plan.md).
  • Inference. The Llama family, Qwen 2 and 3 and their hybrid and MoE successors, Gemma 1 to 4, Phi, Granite, GPT-2, Mistral 3, LFM2, StarCoder2, DeepSeek-Coder, GLM, gpt-oss and others, each published file compared with a reference (models.md, compatibility.md, inference.md, moe.md). MoE experts streamed from disk for models larger than memory.
  • GPU. Vulkan as the main backend and OpenCL as the fallback, on AMD and Intel, every op checked against the CPU kernels; layers split between the device and the CPU (vulkan.md, opencl.md, benchmarks.md).
  • Training. From scratch, fine-tuning, LoRA on a quantized base, reinforcement learning from rewards, on the CPU, OpenCL and ROCm (training.md); refusal analysis and abliteration.
  • Serving. oklm serve: OpenAI-compatible and Anthropic-compatible routes, streams, tool calls, JSON schemas, continuous batching on a paged cache, API keys, TLS, a dashboard, embeddings, reranking, speech both ways, another model loaded in place (server.md).
  • Agent. oklm agent: permission rules and modes, a Linux sandbox, tools, skills, sub-agents, background processes, git, slash commands, MCP, sessions (user-guide.md, extending.md, gap-analysis.md).
  • Retrieval. oklm embed, oklm rag, oklm rerank, retrieval in run, chat, serve and the agent (rag.md).
  • Images, sounds, videos, speech. oklm describe with the LLaVA, SmolVLM, Gemma 3, Qwen2-VL, Qwen2.5-VL and Ultravox projectors, oklm transcribe with Whisper, oklm speak with OuteTTS and Soprano, the encoders and the vocoders on the device (multimodal.md, speech.md).
  • Chats and windows. oklm chat in the terminal with attachments, oklm-gui with a chat window and the graphical tools (models, servers, retrieval, training, inspection, conversion, sessions) (gui.md, commands.md). The server manager and the chat window set the options of oklm serve (layers on the device, the cache's type and size, experts on the CPU, sampling, and the rest in a shared Advanced window), keep them for each model, and say before a server starts what it will take of the memory, as the server sizes it. A chat window started with nothing named loads no model: it lists the servers running on the machine and the installed models.
  • A full context. oklm chat, the chat window, the dashboard's chat and oklm agent summarize the older messages into a note when the context is nearly full (/compact) and go on with a reply the context or the token limit cut (/continue), by themselves or when asked; /v1/chat/completions continues a last assistant message left open (continue_final_message, prefix), in the channels of a model that reasons in one (harmony, Gemma 4, Muse Glimmer) too. A reply cut inside a tool call is completed and never run half written: the answer carries the call's text so far (tool_call_open) in place of a call. POST /api/compact compacts a conversation for a client that cannot. oklm chat and oklm agent share the two settings (autoCompact, autoContinue, continueMax in the user's settings), and the option that goes on with a kept conversation or session is --resume in both (--continue is refused with a line that names it) (inference.md, server.md).

how-it-works.md explains each of these for a newcomer; ROADMAP_DONE.md holds every finished item with its proof.

The models of 0.1.0

OKLM-Nano, OKLM-Micro, OKLM-Micro-Stories and OKLM-Small stay at version 0.1.0 (release-notes-0.1.0.md): this release does not train them again, and 1.0.0 loads them as 0.1.0 did.

What stays compatible

The promise is in the README, "Compatibility across versions": within 1.x the command line, the /v1/ routes, GGUF and safetensors files, OKLM's own versioned files (the KV file) and the agent's settings are kept compatible; the C API, the /api/ routes, the output written for programs, caches and derived files, and results to the last bit may change between minor versions, each change named in the release notes. A command or an option that has to go is announced a minor version before.

Platforms

  • Tested: Linux on x86-64, with an AMD RX 7900 XT and an Intel UHD 750, on Vulkan and OpenCL.
  • Built, tested under Wine only: Windows (MinGW). Not run on a real Windows machine: the job limits, the cmd.exe shell, the microphone, the paste of files and the line editor of oklm chat are in roadmap.md as not run or not written there.
  • Written, never compiled: the macOS branches.
  • Untested: NVIDIA GPUs. tools/gpu/report.sh writes the report an NVIDIA owner can send back (testing.md).

Packages

Built from this tree by make deb, make appimage and make installer-windows into dist/: oklm_1.0.0_amd64.deb, oklm-1.0.0-x86_64.AppImage and oklm-1.0.0-setup.exe (build.md). Each holds oklm, oklm-gui and the HTML documentation. Their SHA-256 sums are in dist/SHA256SUMS-1.0.0. The packages built in a Debian container (make docker-build, named with the release's codename) were not built for this version.

Known limits

The open list is roadmap.md. What a user is most likely to meet:

  • Models larger than the device run with layers or experts on the CPU; several architectures were checked on their smallest published file only, for lack of disk.
  • mimo2 runs on the CPU and was checked on a tiny model and a published header, not on published weights.
  • OpenCL decodes slower than Vulkan on some quantized files; Intel's OpenCL on the UHD 750 is several times slower than its Vulkan.
  • Speech synthesis is English with one built-in voice and speaker files; recognition shows the words while speaking only as fast as the recognizer runs again, which needs a device with the large models.
  • The agent's remaining gaps against its specification are in gap-analysis.md.