mlx-optiq

Changelog

Every version of mlx-optiq published to PyPI, what shipped, and why. The source is the repo's CHANGELOG.md; this page is its public mirror.

v0.8.2

Added

  • nudges setting (default on). Set it to false and an unattended optiq code run leaves stalls, repeated calls and overdue test runs to the model.
  • The local tool takes timeout_seconds.

Changed

  • An unattended optiq code run is no longer asked to review its work before it finishes.
  • The local worker gets at most 40% of an unattended run's time, and never the last 10%.
  • A call repeated with nothing changed gets the "change approach" message at the third refusal, not the twentieth, and the run stops at ten.

Fixed

  • optiq code's run_tests uses the repo's own runner (go test, cargo test, npm test) and counts Go results.
  • A timed-out command is stopped with everything it started.
  • Any file the agent writes counts as a change.
  • run_tests falls back to python -m unittest in a Python repo with no pytest installed.

v0.8.1

Added

  • optiq serve uses the tested speed kernels on Groot-19B-A3B-OptiQ-4bit.

Fixed

  • optiq convert quantizes the routed experts of an MoE model's MTP head when the source stores them fused. They stayed at bf16, so --mtp refused the head.
  • A model with no MTP head logs an info line, not an "injection failed" warning.

v0.8.0

One config, breaking. See Configuration.

Changed

  • All settings live in one YAML file, ~/.optiq/config.yaml, or <repo>/.optiq/config.yaml for a repo. The first run moves your old user files and keeps them as .bak, so you stay logged in. Move a repo's .optiq/code.json or .optiq/optiq.json yourself.
  • optiq config lists everything. optiq code config is gone.
  • Every setting has one variable, OPTIQ_<NAME>: OPTIQ_CODE_MODEL is now OPTIQ_MODEL, and so on. An old or unknown OPTIQ_ variable is ignored with a warning.
  • optiq convert: --target-bpw is --target-bits, --n-calibration is --n-samples, --target is gone.
  • optiq eval: --reference-mode is --reference, --skip-kl is gone.
  • optiq prune-experts: --target-retention is now --retain with a fraction.
  • --stream-experts takes auto, on or off everywhere.
  • optiq lab: --lab-port is --port. optiq code: -y's long form is --auto-approve.
  • OPTIQ_LOWBIT_SEARCH is OPTIQ_LOW_BIT_SEARCH, and OPTIQ_SERVE_ON_GENERATION_DEATH is OPTIQ_ON_GENERATION_DEATH.
  • Codex launched through OptiQ reads its key from OPTIQ_API_KEY, not OPTIQ_AUTH_TOKEN.

Removed

  • 26 internal tuning settings, now fixed: compact_headroom, compact_target, compact_floor_chars, compact_min_yield, compact_keep_recent, explore_limit, edit_fail_limit, dup_call_limit, max_truncations, mcp_defer, warm_cache, line_numbers, edit_tolerance, transport_retry_s, guidance_max_chars, local_tool_max_turns, local_tool_max_seconds, test_nudge_edits, flag_test_edits, server_ready_timeout, server_host, lowbit_search_max_bits, stream_scales_budget_gb, cluster_gpu_reserve_gb, cluster_cache_mb, fused_ce_budget_mb.

v0.7.11

Added

  • OptiQ Code keeps a killed session's conversation, written every turn, so -c resumes it. Each session also logs one line per model call: timing, tokens and status.

Fixed

  • optiq serve --mtp applies the model's sampling penalties as plain serving does. It ignored them, so it could repeat itself and wrote different text. Structured output under --mtp uses plain decoding.
  • optiq serve --mtp reports length when it runs out of tokens, and no longer goes one token over max_tokens.

v0.7.10

Fixed

  • The MTP head now loads on quants that store its experts one by one, such as Qwen3.5-35B-A3B and Ornith-1.5-35B-A3B. It was refused before, so they drafted without it.
  • run_tests in OptiQ Code retries its default as python -m pytest when collection cannot import a module, so a repo with a module at its root no longer reports passing tests as failed.
  • OptiQ Code's final answer is written for the person reading it, rather than repeating its completion review item by item.

v0.7.9

Added

  • Liquid's d1 decision models (LiquidAI/d1-* and their quants): optiq.decision.D1, and optiq serve answers them on /v1/systemone. Yes/no, choice, score and estimate questions over text and JSON states; span reads and pictures not yet.

Fixed

  • LFM3 models no longer abort the server on the first request handled off the loading thread ("There is no Stream(gpu, 0) in current thread").
  • optiq convert no longer fails with "load_tokenizer() got multiple values for argument 'tokenizer_config_extra'" when a model with custom code was loaded earlier in the same process.

Earlier releases

The 14 releases before v0.7.9. Click any version to expand it. Everything before v0.6.3 is in the archive.

v0.7.8

Added

  • LFM3 (Liquid): lfm3 text models and the LFM3-VL vision front-end. Convert, eval, kv-cache, serve and images through the vision sidecar.

Fixed

  • optiq kv-cache and the Python engine load vendored architectures (LFM3, Xing4.0, LLaDA2, Mistral 4); they failed with "Model type ... not supported".
v0.7.7

Added

  • OptiQ Code sandbox, off by default: /sandbox, --sandbox or "sandbox": true runs bash and run_tests under sandbox-exec (macOS) or bwrap (Linux). Writes only to the repo, temp and package caches; credential files unreadable; secret environment variables removed; network only to an allowlist of package registries and git hosts, with a prompt for anything else.

Changed

  • The Lab's python and terminal tools share the same sandbox code: credential files are now unreadable there too, CPU, process and file-size limits apply, and Linux without a container runtime uses bwrap instead of running unconfined.
v0.7.6

Fixed

  • optiq run claude no longer makes Claude Code warn that claude.ai connectors are disabled. They never load under an OptiQ token; locally configured MCP servers work as before.
  • optiq eval --served matches in-process scoring again: it sends penalties as off instead of leaving them out, which let the server apply the model's recommended presence penalty.
v0.7.5

Changed

  • optiq serve passes a requested reasoning effort through to the model: any effort other than none/minimal, or Anthropic thinking.type: enabled, turns thinking on, and the effort reaches templates that read reasoning_effort (Spark-X2.6, gpt-oss). A template that rejects the value (Spark-X2.6 takes only low and high) renders at its default instead of failing the request.
v0.7.4

Added

  • OptiQ Code reads Agent Skills (SKILL.md folders) from .agents/skills and .claude/skills, in the project and your home directory. The agent loads one with the new skill tool when a task matches, /<name> runs one, and /skills lists them.

Fixed

  • OptiQ Code counts tests run through bash (for example python3 -m pytest) like run_tests: the run reports their result instead of "0 passed / 0 failed", and a passing run is no longer nudged as stuck on the next turn.
  • optiq serve honours the standard thinking controls: reasoning_effort (Chat Completions), reasoning.effort (Responses) and thinking.type (Messages). "none", "minimal" or "disabled" turn thinking off; they were ignored, so the model reasoned anyway. An explicit chat_template_kwargs still wins.
v0.7.3

Fixed

  • optiq run claude --cloud no longer makes Claude Code warn that its auto-compact window is too large: it is held to the 1M Claude Code accepts.
  • The Lab reads an OptiQ Cloud model's window from the model list. It asked about the local model's window instead, or got nothing on a cloud-only server, and ran with compaction off and an 8k reply budget.
v0.7.2

Changed

  • optiq run --cloud gives the agent the model's full listed context window (about 1M tokens) instead of capping it at 200,000; each agent decides when to compact.

Fixed

  • Claude Code on an OptiQ Cloud model could print a reply starting with {"type": "thinking", ...}: the Anthropic shim turned earlier thinking blocks into text. They now travel as reasoning, never as text.
v0.7.1

Changed

  • optiq serve refuses a --host other than this machine unless it has an API key (--api-key or OPTIQ_API_KEY), and then every request must carry that key as Authorization: Bearer or x-api-key. On 127.0.0.1 nothing changes.
  • optiq serve --cloud-only and optiq cluster serve follow the same rule.
  • optiq lab --host exposes only the Lab UI; its model API stays on 127.0.0.1.
v0.7.0

Added

  • optiq serve lists OptiQ Cloud's models as optiq-cloud/<id> once signed in and answers them for every client.
  • optiq serve --cloud-only serves only OptiQ Cloud models, with no local model loaded.
  • optiq run <agent> --cloud <model> runs Claude Code, Codex, OpenCode, Mistral Vibe, Hermes Agent or OpenClaw on an OptiQ Cloud model.
  • optiq cloud models lists OptiQ Cloud's models with their prices and context windows.
  • OptiQ Code: /boost <model> and the boost_model setting pick the cloud model Boost runs on; /boost off turns it off.
  • The Lab lists OptiQ Cloud's models, and optiq lab serves them when signed in with no local model.

Changed

  • Claude Code and Codex receive OptiQ Cloud replies as a stream, including tool calls, instead of all at once.

Fixed

  • optiq run hermes failed with Hermes Agent 0.16, which sends its key only to openai.com.
v0.6.7

Changed

  • OptiQ Code Boost now runs through optiq serve on the same Mac (the served model optiq-boost), not by calling OptiQ Cloud directly.
  • Boost needs optiq serve running and logged in with optiq cloud login; without it, Code runs on the local model and says how to turn Boost on.
  • optiq serve streams OptiQ Cloud's reply to OpenAI-format clients as it arrives, instead of buffering it.
  • Boost's local helper returns what it found: no cap on its answer, a time limit, and a final wrap-up turn.
  • Boost passes the cloud model its own earlier reasoning on later turns.

Fixed

  • optiq serve puts back the reasoning that a client dropped from earlier tool-call turns, so the model sees its own reasoning and the prompt cache is reused on hybrid models.
  • Boost continues on the local model when OptiQ Cloud cannot answer, and retries a cut stream.
v0.6.6

Changed

  • Lab hybrid search lists keyword matches first and gives the last places to meaning matches keyword search missed. Document search returns 10 results by default (was 5).
  • Lab deep research follows the Thinking toggle and the thinking budget.

Fixed

  • A Lab reply that ended inside the model's reasoning came back empty. The Lab now asks once for the answer.
  • Rebuilding a meaning index at a new vector size mixed old and new vectors.
v0.6.5

Added

  • Lab and OptiQ Code draft a bug report when they crash. Nothing is sent until you review it and choose Send.
  • The Lab's feedback form asks what kind of feedback it is: something's wrong, an idea, or something else.
v0.6.4

Added

  • Lab collections: upload documents and search them in chat and deep research, by keyword and by meaning.
  • optiq serve --embedding-model serves /v1/embeddings from a Qwen3-Embedding model next to the chat model.
  • MCP 2.x SDK support, alongside 1.x.
  • OptiQ Code temperature and top_p settings for your own endpoints.

Changed

  • Sampling defaults come from each model's generation_config, including presence and frequency penalties. OptiQ Code and the Lab no longer override them.
  • optiq serve checks the port before loading the model.

Fixed

  • Thinking budgets had no effect on mlx-lm 0.32. The Lab now sends one.
  • The Lab read the context window from the wrong route and capped every turn at 8k tokens.
  • Chunked verify never turned on in optiq serve.
  • A failed server start could hang with the model loaded, or restart into the same busy port.
v0.6.3

Changed

  • optiq convert writes the vision towers beside the model shards instead of under optiq/, and copies the base model's image processor files. LM Studio downloads and oMLX now load the vision towers; the MTP head stays in optiq/.

Fixed

  • Chunked verify attention never ran: the GQA decode kernel installs first, and chunking only replaced the unwrapped function. It now runs below 16k context, where 5- and 9-position verify passes are 1.12-1.37x faster, and is on for five REAP-19B models.