Skip to content

Changelog

What changed in the Decision API, the Myles models, the SDKs and the MCP server, and when.

31 October 2026

  • Myles-400M, Myles-150M and Myles-22M are now requested in the console instead of on Hugging Face. We review every request, and approved organisations download the model files and myles.py from the console. See Model access.

29 October 2026

  • Myles-G 12B has a new build, myles-g-12b-20261029, with stronger reasoning and calculation. myles-g-12b now answers with it, so requests to myles-g-12b move to the new build automatically and need no change. The previous build, myles-g-12b-20261017, is retired today: a request naming it is answered 410 with myles-g-12b-20261029 as its successor.

27 October 2026

  • Myles-G 31B has a new build, myles-g-31b-20261027, trained further on tool use and on knowing when it does not know. myles-g-31b now answers with it. The previous build, myles-g-31b-20261024, could not stay on the same hardware, so it is retired today: a request naming it is answered 410 with myles-g-31b-20261027 as its successor. Requests to myles-g-31b need no change.
  • Long states on Myles-G 31B: its context window is now 24,576 tokens, up from 16,384, and a state that does not fit is refused with 422 context_length_exceeded instead of being trimmed, so the model never decides on part of a state. Myles-G 12B is unchanged: it still trims a state too long for its window. See State.
  • On a choice question with exactly one option meaning "I don't know", Myles-G 31B now puts more weight on that option when it is unsure, rather than on a guess.
  • The Mellix MCP server, mellix-mcp@0.1.0 on npm (Apache-2.0): any MCP client can call Myles as a decision tool with npx -y mellix-mcp@0.1.0. It has three tools, decide, list_models and health. Every answer carries its probabilities, whether it cleared your threshold and whether to hand it to a person, and long yes/no checks can be read in overlapping windows with chunk: true. New page: MCP.

24 October 2026

  • The old test gateway address is retired, so do not use it. https://api.mellix.com is the one address for every hosted model, with the same keys. See Hosted.
  • Long states on Myles-G 12B and Myles-G 31B: the model now reads the whole state whenever the request fits its context window, where before it read only the first 20,000 characters. A state too long for the context window is trimmed to fit, keeping its beginning and its end, and the response says so with "truncated": true, state_chars and the X-State-Truncated header. A long state is never refused. Requests whose state is up to 20,000 characters are answered exactly as before. See State.
  • Myles-G 31B now takes the same request sizes as Myles-G 12B: up to 256 questions a request and 256 options a question, 4,096 options over a request. Requests within the old limits are answered exactly as before.

20 October 2026

  • Myles-22M version 4, a new build of the browser model. On release gate v3, which weighs test sets written and labelled by people, it scores 0.9 points (95% interval +0.6 to +1.2) above version 3, and a second training run with a different seed confirmed the gain. The browser demo, the model card and the gated weights on Hugging Face are updated. Version 3 stays available as the revision tagged myles-22m-g12-distill-soup. The CPU models have since moved to the console (see 31 October).

19 October 2026

  • The Batch API: send many decisions as one JSON Lines file to POST /v1/batches and collect the answers within 24 hours, at half the live price per input token (£0.00075 per million on Myles-G 12B, £0.0025 on Myles-G 31B, output tokens free), from the same credit with the same keys. Every line names its model and carries your custom_id, and results come back in input order with usage per line. There are status, results and cancel calls, a list of your batches and an optional webhook. A batch holds up to 50,000 lines or 32 MiB, and a key can have 10 open batches. Batch files are stored encrypted in Canada and deleted: your lines when the batch finishes, and the results 7 days after they are first collected or 14 days after the batch finished, whichever comes first. New page: Batches.
  • Version 0.5.0 of the JavaScript and Python SDKs send batches: client.batch(lines), then wait for the batch and read its results as they arrive. The Python SDK now has a page of its own.
  • Every request example now names its model. Always send model on POST /v1/decide; see Decide.
  • Myles-G 31B on the hosted API for everyone: model id myles-g-31b at https://api.mellix.com, £0.005 per million input tokens with output tokens free (Myles-G 12B stays at £0.0015), paid from the same prepaid credit with the same keys. Text and images, 16,384 tokens of context. Requests that name myles-g-31b-preview keep working and are answered by myles-g-31b. See Hosted.
  • One address: https://api.mellix.com now serves every hosted model, Myles-G 31B included. Name the model in every request, "model": "myles-g-12b" or "model": "myles-g-31b", with the same keys. See Hosted.

17 October 2026

  • Yes/no answers are committed on both Myles-G models: probabilities for uncertain answers are reported as at least 0.80 towards the chosen side (0.801 today), so noul is never between 0.2 and 0.8. Other answers are unchanged. See Question types and Thresholds.
  • Images on Myles-G 31B (then myles-g-31b-preview): it reads PNG and JPEG images in assets exactly as Myles-G 12B does, with the same limits, image_budget and billing. Its context window is 16,384 tokens. See Images.

16 October 2026

  • choice and score answers carry confidence, and score answers carry legend. See Confidence.
  • noul criteria with null descriptions are read as Yes and No. Criteria that name only one side, or use keys other than true/yes and false/no, now return 422 instead of an answer.
  • Images on the Decision API: Myles-G 12B reads PNG and JPEG images sent in assets on POST /v1/decide, referenced by name from the state, the instructions or the criteria descriptions. A new image_budget sets how many tokens one image may use. Images are billed as input tokens, and an image in the state once per request. New page: Images.
  • A request with images for a model that does not read them (any model other than Myles-G 12B) is refused with 422 images_not_supported, never answered without its images.
  • Request bodies up to 32 MiB, gzip and zstd compression, and multipart/form-data bodies (415 for other encodings).
  • X-Request-Id on every response, Retry-After on 429 and 503, field paths in 422 errors, and models beside data in GET /v1/models.

15 October 2026

  • New page, Model types: Myles-G is called through the Decision API with POST /v1/decide, and the CPU models run in your own program with decide(state, question, options). No Myles model generates text.

14 October 2026

  • New documentation site at /docs: guides, API reference, model library and deployment.

11 October 2026

  • Decision API version 1: POST /v1/decide with noul, choice and score questions, GET /v1/models and GET /health.
  • POST /v1/systemone for compatibility with existing decision-model clients.