Skip to content

MINIMAX_M2_5_JANG_MACBOOK2.md

Latest commit

 

History

History
97 lines (61 loc) · 2.86 KB

MINIMAX_M2_5_JANG_MACBOOK2.md

File metadata and controls

97 lines (61 loc) · 2.86 KB
Raw

MiniMax-M2.5 JANG_2L on macbook2

Date: 2026-03-28

Bottom Line

JANGQ-AI/MiniMax-M2.5-JANG_2L does load and generate on macbook2 with the existing ~/python3-venv/torch313 environment.

The missing piece was not a separate MLX fork. The required runtime is the jang package layered on top of MLX:

  • jang 2.2.0
  • mlx 0.30.6
  • mlx-lm 0.30.7

That stack was already present in the venv and was sufficient to load the cached local model snapshot.

Sources

From those sources:

  • temp=1.0 is required; greedy temp=0 can loop
  • top_p=0.95, top_k=40 are recommended
  • JANG models need the JANG runtime rather than plain MLX loading

Local Model Path

Used local snapshot path only, with no new download:

/Users/ljubomir/.cache/huggingface/hub/models--JANGQ-AI--MiniMax-M2.5-JANG_2L/snapshots/6a2d1e1c9475a63961cfb17253436dbe5fd6fa65

The models/huggingface/... path resolves to the same cache directory.

Model Facts From Local Config

  • model_type=minimax_m2
  • max_position_embeddings=196608
  • num_hidden_layers=62
  • num_local_experts=256
  • num_experts_per_tok=8
  • jang_config.runtime.total_weight_gb=55.9
  • Disk footprint on this machine: about 63 GB

What Was Run

Offline local-path run in ~/python3-venv/torch313:

source ~/python3-venv/torch313/bin/activate
export HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1
python scripts/run_minimax_jang_example.py --repeat 2

The helper script is here:

Measured Results

Cold-load run:

  • load time: about 19.8s to 31.2s
  • first token on first short prompt after load: 17.75s
  • decode speed: about 0.52 tok/s

Two prompts in one loaded process:

  • run 1: first token 30.64s, decode 0.51 tok/s
  • run 2: first token 5.86s, decode 0.51 tok/s

So the practical picture is:

  • the model fits
  • the loader works
  • warm TTFT gets much better than cold TTFT
  • steady decode remains very slow on macbook2, around 0.5 tok/s

Memory

Observed in the short measured run:

  • peak memory footprint: 67894839232 bytes, about 67.9 GB
  • current iogpu.wired_limit_mb on this machine was 0

So the example fit without raising the wired GPU limit to 88000.

Output Quality Note

The raw example prompt from the model card does generate, but it tended to continue in an odd templated style rather than producing a crisp one-line answer immediately. That does not change the runtime result: the core runtime stack works, but the model is not practically fast on this hardware in this configuration.

Conclusion

This MiniMax JANG_2L setup is technically runnable on macbook2, but speed is not good enough to be attractive for interactive local use.