DEV.co
Open-Source LLM · XiaomiMiMo

MiMo-V2.5-Pro-FP4-DFlash

MiMo-V2.5-Pro-FP4-DFlash is a 1-trillion-parameter mixture-of-experts language model from Xiaomi optimized for fast inference. It uses FP4 quantization on its expert layers and a block-diffusion speculative decoding technique (DFlash) to reduce both model size and decoding latency. The model supports up to 1M context length and is available on HuggingFace under an MIT license without gating.

Source: HuggingFace — huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash
554.3B
Parameters
mit
License (OSI-approved)
Unknown
Context (tokens)
47k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
DeveloperXiaomiMiMo
Parameters554.3B
Context windowUnknown
Licensemit — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads47k
Likes142
Last updated2026-06-08
SourceXiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash

What MiMo-V2.5-Pro-FP4-DFlash is

A trillion-scale MoE transformer (70 layers, 128 attention heads, 6144 hidden size) with selective FP4 quantization applied only to MoE experts while keeping attention projections at higher precision. Paired with a 5-layer BF16 DFlash drafter that performs block-level masked parallel prediction (8-token blocks) to enable speculative decoding with ~4–6 tokens acceptance length. Supports 1M context via sliding-window attention (128-token window in backbone, 1024 in drafter). Deployment via SGLang with tensor parallelism (e.g., 16-way) and expert parallelism on compatible hardware.

Quickstart

Run MiMo-V2.5-Pro-FP4-DFlash locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

High-throughput code and software engineering

SWE-Bench Pro score of 58.8% and per-token compute moved to constant time with DFlash makes this suitable for large-scale code completion and bug-fixing workflows where latency and throughput are critical.

Long-context retrieval-augmented generation (RAG)

1M token context with efficient sliding-window attention enables RAG pipelines over large documents without context fragmentation, particularly valuable for legal, technical, or financial document analysis.

High-concurrency API serving at scale

FP4 quantization and DFlash speculative decoding reduce per-request memory bandwidth and forward-pass count; suitable for production inference services requiring thousands of concurrent requests on large GPU clusters.

Running & fine-tuning it

Requires multi-GPU, multi-node setup: minimum estimated 256–512 GB total GPU VRAM for distributed inference (estimate based on 1T params in FP4/mixed precision with tensor parallelism). SGLang example uses 16-way tensor parallelism and 2-way data parallelism (32 GPUs suggested). Interconnect bandwidth (e.g., NVLink) strongly recommended for expert-parallel communication. Exact VRAM footprint depends on context length, batch size, and parallelism strategy; verify on your hardware.

No explicit LoRA or QLoRA guidance in card. FP4 quantization applied to experts post-training; standard LoRA over quantized weights may degrade performance. Full fine-tuning on FP4 backbone is undocumented and requires re-quantization validation. Recommend contacting [email protected] for fine-tuning best practices or use the Xiaomi MiMo API platform if custom adaptation is needed.

When to avoid it — and what to weigh

  • Single-GPU or constrained memory deployments — Even with FP4 quantization, 1T parameters require multi-GPU, multi-node setup with expert parallelism. Not feasible on consumer hardware; requires data-center scale infrastructure.
  • Low-latency, latency-sensitive applications — While DFlash improves throughput, per-token latency remains governed by backbone forward-pass cost. Real-time interactive systems (e.g., sub-100ms response) should evaluate empirically first.
  • Fine-tuning or continued pretraining — FP4 QAT was applied post-hoc; further fine-tuning on FP4 backbone is not documented. LoRA on quantized experts or full-parameter retraining would require careful re-quantization and validation.
  • Non-English or specialized domain tasks without adaptation — Model trained on general agent and code tasks. Performance on domain-specific corpora (biomedical, legal, specialized technical) not benchmarked; may require fine-tuning or prompt engineering validation.

License & commercial use

MIT license. MIT is a permissive OSI-approved license allowing commercial use, modification, and distribution with minimal restrictions (attribution required).

MIT license permits commercial use, provided attribution to the original work is included. No gating or access restrictions on the model download. However, deployment at trillion-parameter scale incurs significant infrastructure cost; evaluate whether Xiaomi MiMo API Platform managed service is more cost-effective than self-hosting.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityHigh
DEV.co fitGood
Assessment confidenceHigh
Security considerations

No explicit security audit or adversarial robustness claims stated. Trillion-scale models trained on broad internet corpora inherit inherent risks: potential for prompt injection, jailbreaks, and bias in code generation and reasoning tasks. Speculative decoding (DFlash) does not alter the backbone's output; verification step preserves output fidelity but does not add security filtering. Quantization (FP4) reduces model size but does not provide content security guarantees. Recommend input validation, output monitoring, and safety testing before production deployment in sensitive domains.

Alternatives to consider

Meta Llama 3.1 405B (or 70B variants)

Openly available trillion-scale model with strong code/reasoning benchmarks. Smaller context (8K base), different quantization strategy (INT8/FP8 common), and broader ecosystem support; consider if you lack trillion-scale infrastructure budget or need proven stability.

Mistral Large 2 (or MoE variant Mixtral 8x22B)

MoE architecture with lower parameter count (~50–100B active); easier to deploy on fewer GPUs. No speculative decoding built-in but simpler tooling (vLLM, TGI). Trade inference speed for reduced hardware complexity.

Qwen2.5-Turbo or Claude 3.5 Sonnet (API)

If self-hosted trillion-parameter serving is prohibitive, managed APIs reduce operational overhead. Evaluate cost-per-query vs. MiMo API Platform pricing and latency SLAs.

Software development agency

Ship MiMo-V2.5-Pro-FP4-DFlash with senior software developers

MiMo-V2.5-Pro-FP4-DFlash offers production-grade code and reasoning at scale. Verify your GPU cluster capacity, test on sample workloads, and contact Xiaomi MiMo ([email protected]) for support. Alternatively, evaluate the Xiaomi MiMo API Platform for managed inference.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

MiMo-V2.5-Pro-FP4-DFlash FAQ

Can I use this model commercially?
Yes. MIT license permits commercial use with attribution. Ensure you include the original copyright notice and license text in your distribution/deployment documentation.
What GPU hardware do I need to run this?
Minimum 32 high-end GPUs (e.g., H100/A100) with NVLink interconnect for expert parallelism. Single-node deployment is not practical. Use the SGLang example as a starting point, adjust --tensor-parallel-size and --expert-parallel-size to your cluster.
Does FP4 quantization hurt accuracy?
Model card shows near-lossless quality on general and code benchmarks (Claw-Eval +6.27%, SWE-Bench +2.80%). However, expert-only FP4 (keeping attention at higher precision) was chosen to minimize degradation. Test on your specific tasks before production use.
Can I fine-tune this model?
Not explicitly documented. FP4 quantization was applied post-training; standard LoRA on quantized weights may not be optimal. Contact [email protected] or explore Xiaomi MiMo API Platform custom adaptation options.

Software developers & web developers for hire

Need help beyond evaluating MiMo-V2.5-Pro-FP4-DFlash? DEV.co is a software development agency offering software development services and web development for teams of every size. Our software developers and web developers build custom software, web applications, APIs, and open-source llms integrations — and maintain them long-term.

Ready to Deploy a Trillion-Parameter Model?

MiMo-V2.5-Pro-FP4-DFlash offers production-grade code and reasoning at scale. Verify your GPU cluster capacity, test on sample workloads, and contact Xiaomi MiMo ([email protected]) for support. Alternatively, evaluate the Xiaomi MiMo API Platform for managed inference.