MiMo-V2.6-Distill-Qwen-9B-FP8

MiMo-V2.6-Distill-Qwen-9B-FP8 is a FP8 dynamic quantized version of XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B, a 9B agentic model developed by Xiaomi MiMo through supervised fine-tuning of Qwen/Qwen3.5-9B on MiMo-generated data. The base checkpoint covers coding, general-purpose agent tasks, visual coding, and cybersecurity, and is released as a starting point for open research in agentic reinforcement learning. This quantization reduces model size and memory footprint while preserving the model's agentic tool use, coding, long-form reasoning, and instruction-following capabilities, making deployment more accessible on smaller GPUs. The checkpoint includes its tokenizer and the MiMo v2.6 chat template, with thinking mode toggleable via enable_thinking.

Quantization to FP8 may introduce minor numerical differences relative to the bf16 source model. This is an SFT-only checkpoint (not RL-trained); behavior may differ from downstream RL-tuned MiMo-V2.6 releases.

Quantization Details

Quantization was performed using llmcompressor with the following recipe:

default_stage:
  default_modifiers:
    QuantizationModifier:
      targets: [Linear]
      ignore: ['re:.*lm_head', 're:.*embed_tokens$', 're:.*visual.*',
        're:.*model.visual.*', 're:.*linear_attn.*']
      scheme: FP8_DYNAMIC
      bypass_divisibility_checks: false
      requires_calibration_data: false

Linear layers are quantized to FP8 with dynamic per-tensor activation scaling, so no calibration dataset is required (requires_calibration_data: false). The lm_head, embedding table, any vision-tower (visual) components, and linear_attn layers are excluded from quantization and remain at full precision to preserve output-head fidelity and numerical stability.

Base model	XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B
Original backbone	Qwen/Qwen3.5-9B
Quantization scheme	FP8_DYNAMIC (Linear layers only)
Format	compressed-tensors
Calibration data required	No (dynamic activation scaling)
Excluded from quantization	lm_head, embed_tokens, visual (if present), linear_attn

Use with vLLM

MiMo-V2.6-Distill-Qwen-9B-FP8 is served through vLLM with native support for compressed-tensors FP8 checkpoints.

Requirements

• torch >= 2.11.0
• vllm >= 0.19.1
• A GPU with FP8 support recommended (Hopper or Blackwell class) for best throughput; also runs on Ampere with FP8 dequantized on the fly.

Serve

vllm serve prithivMLmods/MiMo-V2.6-Distill-Qwen-9B-FP8 \
  --max-model-len 32768

Client request

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

messages = [
    {
        "role": "user",
        "content": "What is 15% of 240?"
    }
]

response = client.chat.completions.create(
    model="prithivMLmods/MiMo-V2.6-Distill-Qwen-9B-FP8",
    messages=messages,
    temperature=0.0,
    max_tokens=2048,
    extra_body={"chat_template_kwargs": {"enable_thinking": True}},
)

message = response.choices[0].message
print("Thinking:", getattr(message, "reasoning_content", "") or "")
print("Answer:", message.content or "")

Quick Start with Transformers

pip install transformers accelerate

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model = AutoModelForCausalLM.from_pretrained(
    "prithivMLmods/MiMo-V2.6-Distill-Qwen-9B-FP8",
    torch_dtype="auto",
    device_map="auto"
)

tokenizer = AutoTokenizer.from_pretrained(
    "prithivMLmods/MiMo-V2.6-Distill-Qwen-9B-FP8"
)

messages = [
    {
        "role": "user",
        "content": "What is 15% of 240?"
    }
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True
)

inputs = tokenizer(text, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=2048
)

print(
    tokenizer.decode(
        outputs[0][inputs["input_ids"].shape[-1]:],
        skip_special_tokens=True
    )
)

Evaluation (Base Model)

Results for the released SFT checkpoint (bf16 source), as reported in the MiMo-V2.6 technical report. The FP8 quantized checkpoint is expected to track these results closely; minor numerical differences from quantization should be validated for production use cases.

Domain	Benchmark	Metric	Qwen3.5-9B	MiMo-V2.6-Distill-Qwen-9B (SFT)
Code	SWE Verified	avg@3	60.0	61.1
Code	SWE Pro	avg@3	32.0	44.6
Code	MiMo Code (mini)†	avg@3	19.5	51.6
Cyber	MiMo Cyber (mini)†	avg@3	5.7	31.3
General	AutomationBench v1.0.6	avg@1	5.0	30.3
General	Terminal Bench 2.1	avg@1	27.0	37.1
General	Toolathlon-Verified	avg@1	25.9	35.2
General	OfficeQA	avg@1	9.0	19.5
General	JobBench	avg@1	2.6	18.3
General	MiMo General (mini)†	avg@1	28.5	62.2
Visual	MiMo Visual Coding (mini)†	avg@1	61.7	64.0

Training Details (Base Model)

Setting	Value
Base Model	Qwen/Qwen3.5-9B
Developer	Xiaomi MiMo Team
Training Method	Supervised Fine-Tuning (SFT) on MiMo-generated data
Total SFT Tokens	77.4B (27.2B loss-bearing)
Domains	Code, Cyber, General, Visual
Chat Template	MiMo v2.6 (thinking toggle via enable_thinking)
Purpose	Starting point for open agentic reinforcement learning research

SFT Data Mixture

Domain	Total tokens (B)	Token share (%)	Loss-bearing tokens (B)
Code	23.2	29.9	7.3
Cyber	11.0	14.2	4.8
General	22.0	28.5	5.7
Visual	21.2	27.4	9.4
Total	77.4	100.0	27.2

Intended Use and Limitations

Intended use, known limitations, training data composition, and responsible use guidance are unchanged from the base model. See the MiMo-V2.6-Distill-Qwen-9B model card for full details.

• Agentic RL Research: Open SFT starting point for agentic reinforcement learning experiments.
• Coding Assistance / SWE Tasks: Repository-level coding and software engineering workflows.
• Tool Use & General Agents: Multi-step tool calling, terminal, and automation tasks.
• Visual Coding: Front-end and UI-oriented visual coding generation.
• Cybersecurity Research: Structured security analysis and problem solving.
• Efficient Local Deployment: Reduced memory footprint enables 9B agentic inference on smaller GPUs.

Limitations

• Experimental Model: Behavior may differ from the base model in certain scenarios.
• SFT-only Checkpoint: This is a pre-RL SFT release; downstream RL-tuned MiMo models may exhibit stronger agentic performance.
• Reasoning Artifacts: Complex reasoning chains may occasionally produce incorrect intermediate steps or conclusions.
• Internal Benchmarks: Some reported evaluation sets (†) are internal and not publicly reproducible.
• Quantization Drift: FP8 dynamic quantization may introduce minor numerical differences relative to the bf16 source model; downstream accuracy should be validated for production use cases.

License

Released under the license terms consistent with the base model. Refer to XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B for the authoritative licensing information.

Acknowledgements

• XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B: Source SFT checkpoint used for this quantization.
• Qwen/Qwen3.5-9B: Original backbone model.
• Xiaomi MiMo Team: MiMo-V2.6 technical report and evaluation results.
• llmcompressor: Used to produce the FP8 dynamic quantization for this release.
• vLLM: Inference and serving.

Citation

If you use this model, please cite the original MiMo-V2.6 release:

@misc{mimo2026v26,
  title={MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement},
  author={{Xiaomi MiMo Team}},
  year={2026},
  howpublished={\url{https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL}},
}