Kimi-Linear-48B-A3B-Instruct-MXFP4-GS32-MLX

<div align="center"> <h1><em><strong>Kimi-Linear-48B-A3B-Instruct · MXFP4 · 32-Group (MLX)</strong></em></h1> <strong>moonshotai/Kimi-Linear-48B-A3B-Instruct</strong> base model that has been quantized<br> with <code>mlx-lm 0.28.4</code> into an MXFP4 format using group size 32 (with selective 8-bit gates for MoE stability).<br> The weights and auxiliary files live in this folder and are compatible with the MLX Runtime on Apple Silicon.

한국어 | English

</div>

Model Summary

• Architecture: KimiLinear MoE decoder-only transformer (27 layers, hidden size 2304, 32 attention heads, 256 experts, 8 experts/token) as defined in config.json.
• Context length: Configured for ~1M-token windows via linear attention blocks (effective window depends on runtime memory and the KV budget you set with --max-kv-size).
• Tokenizer: tiktoken-based BPE (tokenizer_config.json, tiktoken.model) with special tokens defined inside the tokenizer files instead of hard-coded IDs.
• Chat template: See chat_template.jinja for the multi-turn schema that mirrors the official Kimi tool-call format.
• License: MIT, matching the upstream moonshotai/Kimi-Linear-48B-A3B-Instruct release.

Quantization Details

• Tooling: mlx-lm (>=0.28.4) via python3 -m mlx_lm.convert -q targeting MXFP4 weights.
• Format: MXFP4 4-bit packing with group size 32 across major linear layers (mode-enforced requirement for MXFP4).
• Exceptions: Mixture-of-experts gate projections remain at 8-bit / group size 64 for routing stability, as recorded in quantization_config.
• Shard layout: 5× model-0000n-of-00005.safetensors plus model.safetensors.index.json for streaming loads.
• Memory: Expect roughly 26–29 GB of Apple Silicon unified memory for the weights; KV cache usage scales with context length.

Quantization snippet (abbreviated from config.json):

"quantization_config": {
  "group_size": 32,
  "bits": 4,
  "mode": "mxfp4",
  "model.layers.1.mlp.gate": {"group_size": 64, "bits": 8},
  "model.layers.2.mlp.gate": {"group_size": 64, "bits": 8},
  "model.layers.3.mlp.gate": {"group_size": 64, "bits": 8},
  "...": "additional gate entries continue through layer 26"
}

Files Included

Intended Use & Limitations

• Use cases: Multilingual assistant/chat, tool invocation, long-context retrieval-augmented generation on Apple Silicon hardware.
• Not for: Decisions requiring guarantees (medical, legal, financial) without human review; handling unfiltered harmful instructions.
• Safety: Mirrors the base model’s safety profile—apply additional filtering and RLHF layers if deploying to end users.
• Security: modeling_kimi.py defines custom modules, so you must use --trust-remote-code with the CLI and trust_remote_code=True with the Python API. Handle sensitive data in an offline, isolated environment.

How to Use (MLX)

1. Install MLX tooling (macOS 13.6+ on Apple Silicon):

   pip install -U mlx-lm  # or: pip install -U "git+https://github.com/ml-explore/mlx-lm.git@main"
   # Offline cache only:
   # HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 ...
   

2. Chat CLI (template + stop rules auto-applied):

   mlx_lm.chat \
     --model /path/to/Kimi-Linear-48B-A3B-Instruct-MXFP4-GS32-MLX \
     --trust-remote-code \
     --max-tokens 512 --temperature 0.7 --top-p 0.9
   

   For ≥256K contexts, consider --max-kv-size 262144 (and scale further as needed).

3. Programmatic usage:

   from mlx_lm import load, generate
   from mlx_lm.sample_utils import make_sampler, make_logits_processors
   
    model, tok = load(
       "/path/to/Kimi-Linear-48B-A3B-Instruct-MXFP4-GS32-MLX",
       trust_remote_code=True,
    )
   
   messages = [{"role": "user", "content": "Kimi Linear 구조를 간단히 요약해줘."}]
   prompt = tok.apply_chat_template(messages, add_generation_prompt=True)
   
   sampler = make_sampler(temperature=0.7, top_p=0.9)
   procs = make_logits_processors(repetition_penalty=1.1, repetition_context_size=64)
   
   print(
       generate(
           model,
           tok,
           prompt,
           max_tokens=512,
           sampler=sampler,
           logits_processors=procs,
       )
   )
   

   (Set HF_HUB_OFFLINE=1 and/or TRANSFORMERS_OFFLINE=1 before running if you need hub-less operation.)

Conversion Notes

• Source checkpoint: moonshotai/Kimi-Linear-48B-A3B-Instruct (synced 2025-11-07 UTC).
• Conversion command actually executed:

  python3 -m mlx_lm.convert \
    --hf-path moonshotai/Kimi-Linear-48B-A3B-Instruct \
    --q-bits 4 -q \
    --group-size 32 \
    -o Kimi-Linear-48B-A3B-Instruct-MXFP4-GS32-MLX
  

• Post-quantization integrity verified via mlx_lm.chat sanity prompts and safetensors checksum inspection.

Integrity & Verification

After upload, verify shard integrity locally:

cd /path/to/Kimi-Linear-48B-A3B-Instruct-MXFP4-GS32-MLX
shasum -a 256 model-*.safetensors > SHA256SUMS
shasum -c SHA256SUMS

Additional Tips

• Set --trust-remote-code whenever loading this repository through MLX CLIs to ensure the custom layers in modeling_kimi.py are registered.
• Leverage prompt caching (mlx_lm.cache_prompt) plus --max-kv-size for reliable ≥1M-token experiments without exhausting unified memory.

Acknowledgments

• Moonshot AI — Thank you for releasing the Kimi family and openly documenting the Kimi Linear architecture. Reference: Moonshot AI GitHub, Kimi Linear repo, and the official technical report.
• Apple Machine Learning Research — Deep gratitude for continuously evolving MLX / MLX-LM so Apple Silicon users can keep learning and iterating. See MLX and MLX-LM.
• MLX Community on Hugging Face — Thanks for sharing MLX-ready weights and examples at lightning speed; they directly inspired this conversion flow. See mlx-community.

Citation

If you use this model, please cite both Moonshot AI and this quantized release in your research or product documentation.