OKLMdocumentation v1.0.3
On this page

OKLM 1.0.1 release notes

What changed between 1.0.0 (2026-11-16) and 1.0.1: the leftovers of the graphical tools' options and of the full context, decided point by point, and the defects met on the way. What 1.0.0 holds, the platforms and the promise of compatibility are in release-notes-1.0.0.md; they stand.

Status: released on 2026-11-17, tag v1.0.1.

The number says fixes, and most of this is; some of it is new (options of the chat window, members of answers, a count in POST /api/compact). All of it is added: nothing of the command line, of the /v1/ routes or of the settings files was removed or changed its meaning, but for the three changes of behaviour named below.

Changes of behaviour

These three change what an existing command does. Each was decided as the lesser harm; the rule that such a change is announced a minor version before (README) was set aside for them, this once, a day after 1.0.0.

  • oklm agent -p ends with status 1 when the message was not answered for want of room (it ended with 0): the agent's system prompt and tools leave less than a turn's room in the context by the model's count, or the model or the server still refuses the prompt for its length after a larger context and a summary were tried. An answer that was written is status 0, one cut by --max-tokens too. The JSON output forms already ended with 1 (agent-sessions.md).
  • oklm agent --url on an oklm serve takes the server's context as its window when the settings' contextWindow does not tell one (it assumed a size of its own). A server's refusal of a prompt for its length is now read as a full context, followed by a larger context where the server resizes its own, a summary where one frees room, or a message that says there is no room with the figures and the ways out (agent-sessions.md).
  • oklm serve with neither --gpu-layers nor --kv-mem, at a context that leaves the device less than one context of cache beside the whole model, places fewer layers on the device (it started whole with a cache shorter than the context): the most layers that leave the device one context of their cache, the others and their cache in host memory, slower than the whole model on the device. Either option given keeps its meaning. A model larger than its device has the same rule in place of a cache of one block (memory-plan.md).

The servers and the memory

  • A model switch never ends with nothing loaded. POST /api/model and the chat window plan the new model before anything is freed; a switch that cannot fit is refused with nothing touched. When memory holds both models the new one is loaded first ("loaded_first" in the answer); otherwise the old one is freed first and loaded again if the new one fails, the error saying so (server.md).
  • The memory line counts what it left out: a projector's working memory on the CPU, a model beside the model on another device (judged on that device, which the line names when it is short), the models the chat window finds by itself for a long document, the speech model with a vocoder's longest call, the expert cache of a MoE model read from disk. GET /api/stats gains host_memory.speak, speak_call, vision, audio, and side_devices; each part of the line is checked against a figure of the server's (gui.md).
  • Fixed: on Vulkan a server sized its cache before the driver had given back the memory of the model it had just freed: after a switch or a resize the cache could be one block and the reply empty.

The chat window and the graphical tools

  • A chat window started with nothing named loads the model chosen in the same window, once the memory line has judged it (a model that does not fit is refused with the reason), and keeps what was typed. Its list has the saved servers, the listening agents and a row to choose any GGUF file, beside the servers of the machine and the installed models (gui.md).
  • The chat window takes the options without a control (an Other tab in its Advanced window, kept for each model) and the MoE disk options (--expert-cache, --pin-gb, --expert-pack).
  • The Advanced window is 740 pixels wide: it fits the smallest window of the tools.
  • Fixed, reading aloud: after a compaction nothing of the reply's second piece was read; a piece asked for with Continue was never read.
  • A compaction for a server at an address counts by the server's tokenizer (count in POST /api/compact), by an estimate only where the server has no such count.

A full context

  • /v1/messages completes a tool call the answer was cut inside (tool_call_open, as /v1/chat/completions does) (server.md).
  • A reply of a model that reasons in a channel, cut inside the first header of its turn, comes back empty in place of a piece of markup and is gone on with from a turn opened afresh.
  • The agent makes its context larger to go on with an answer, and before a turn a conversation no longer fits, where the memory holds it or the server resizes its own; it summarizes otherwise, and makes no summary that frees nothing.
  • The dashboard's chat makes the context larger before it compacts; it was run in Chrome, Firefox and WebKitGTK.
  • oklm chat's full-screen display shows a continued reply as one block.
  • Fixed: oklm agent ended on a double free when a summary was made under an answer being gone on with whose reasoning was kept; an agent on a server sent an answer back without its reasoning block, and an answer cut inside its reasoning never went on.

Checks before the tag

The sanitizer and Wine runs of 1.0.0 are in roadmap.md. For 1.0.1 the whole suite, the graphical tools' checks on the CPU and on the device, the static analysis and the documentation's lint ran on the tree of the tag less its last commit (a test's call the Windows C library does not have, and this status), and the test programs changed since 1.0.0 ran again under ThreadSanitizer, AddressSanitizer and Wine on the tagged tree. The kept servers' check ran against systemd's user manager.

Packages

Built from this tree by make deb, make appimage and make installer-windows into dist/: oklm_1.0.1_amd64.deb, oklm-1.0.1-x86_64.AppImage and oklm-1.0.1-setup.exe, their SHA-256 sums in dist/SHA256SUMS-1.0.1. The packages built in a Debian container (make docker-build) were not built for this version.

Known limits

Those of 1.0.0 stand. What these changes leave is in roadmap.md under T11 and C10; what a user is most likely to meet:

  • A running server asked for a larger context does not move layers off the device: the split is made at start.
  • A reply cut inside a tool call is not completed for a model whose calls no marker opens (Llama 3's bare JSON, GLM-4), nor with oklm agent --no-tool-grammar.
  • gpt-oss and Muse Glimmer continue a reply through a server on tiny models only; Gemma 4 was run on a published file.
  • The first request of an agent's conversation is sent before the model has counted the prompt: in a context too small it costs one request before the message that says so.