MacJev-322M-4K-Laya-MLX

MacJev-322M-4K-Laya-MLX

Website: jevstyle.com — all JevStyle decision models, benchmarks and quickstart in one place.

MacJev running natively on Apple silicon with MLX. No PyTorch, exact FP32 results.

MacJev-322M-4K-Laya is a compact decision model for local Mac agents. You give it an observed state and one typed question with candidate answers. In a single forward pass it returns a probability for every candidate. Use it to rank candidate actions, route tool calls and check task state.

This repository contains the complete model and macjev_mlx.py, a native MLX runtime for all of it: the mmBERT/ModernBERT encoder, type embeddings, decision transformer, candidate scorer and tokenizer.

Highlights

Compared with the Laya multilingual checkpoint it was trained from, with the same 4096-token input budget, on held-out test sets:

• Reads long observations. Yes/no task-state checks on 2K–4K-token inputs rise from 31% to 89% accuracy.
• Probabilities you can threshold. Calibration error on long inputs is about 8 times lower, so stated confidence closely tracks actual accuracy.
• Better on public benchmarks too. Across 11,600 public decisions from typed decisions, Emotion and AG News, accuracy rises by 1.5 points, with a 95% interval of 1.1 to 1.8.
• Reproducible. The long-input and public-benchmark gains appear in all three independently trained seeds. This release is the seed chosen in advance on development data.

Exact and fast on a Mac

• Same answers as the reference. On all 9,000 validation decisions, the MLX runtime picks the same top answer as the PyTorch FP32 model, with probabilities within 0.00026.
• Light dependencies. Only mlx and tokenizers; no PyTorch or Transformers.
• Checked on load. SHA-256 checksums of every file it loads are verified in about 2 seconds.

Forward-pass time on an M1 Max with MLX 0.32.2:

Quickstart

Requirements: Apple silicon Mac, Python 3.11 or newer.

hf download chaoliangUNSW/MacJev-322M-4K-Laya-MLX --local-dir MacJev-MLX
cd MacJev-MLX
python -m pip install -r requirements.txt

from macjev_mlx import MacJevMLX

model = MacJevMLX(".")  # verifies checksums, loads the FP32 weights
result = model.decide(
    state={
        "request": "打开浏览器",
        "available_actions": ["open_browser", "copy_file", "ask_user"],
    },
    question={
        "t": "choice",
        "ins": "Which available action best matches the user's request?",
        "crit": {
            "open_browser": "Open the user's browser",
            "copy_file": "Copy a file inside the approved folder",
            "ask_user": "Ask for clarification",
        },
    },
)
print(result["answer"], result["probabilities"])  # A decision only. Nothing is executed.

Or download from Python, fetching only the files it needs:

model = MacJevMLX.from_pretrained("chaoliangUNSW/MacJev-322M-4K-Laya-MLX")

The result contains answer, probabilities, top_probability, input_tokens, latency_ms, and actions_executed: False. Keep one model object loaded across requests.

Other question shapes:

score = {"t": "score", "ins": "How complete is the observed task?",
         "crit": ["Not started", "Partly complete", "Complete and verified"]}
boolean = {"t": "noul", "ins": "Does the observation prove that the requested file exists?"}

Score answers are ordered string indices ("0", "1", ...). Noul answers are "false" and "true". Choice keys keep their insertion order.

Command line, one JSON object per line with state and question:

echo '{"state": {"request": "open the browser"}, "question": {"t": "noul", "ins": "Does the user want a browser opened?"}}' \
  | python macjev_mlx.py --model .

Inputs can use up to 4096 tokens in total and 1024 for the question and options. Anything larger raises InputBudgetError instead of being truncated, so no candidate or evidence is dropped without you knowing.

Using it with LM Studio

MacJev is a decision model rather than a chat model, so it runs through macjev_mlx.py instead of LM Studio's chat engine. To let an LM Studio chat model use it, expose decide() as a tool or MCP server and have the chat model call it.

Files

Other formats

• PyTorch FP32, with training details
• GGUF for llama.cpp

License and attribution

Apache-2.0, inherited from Laya; see LICENSE and NOTICE. The mmBERT-base backbone is MIT-licensed; see MMBERT_LICENSE. MacJev is an independent project. It is not affiliated with or endorsed by JEV or the Laya authors, and it contains no JEV weights.

• Laya model and source
• mmBERT-base

Community

Interested in compact decision models, local agents, and practical macOS automation? Join our Discord community.

如果你也对轻量决策模型、本地 Agent 和 macOS 自动化感兴趣，欢迎加入我们的 Discord 社区。