MiMo-V2.5-Pro (Xiaomi 1T MoE)

Deploy MiMo-V2.5-Pro (1.02T MoE, 42B active, 1M context) by Xiaomi on Clore.ai — the first open-weight Pro tier from the MiMo team, FP8 native, hybrid attention

Status (April 2026): MiMo-V2.5-Pro was released on April 27, 2026 by Xiaomi's AI division as the first open-weight model in their Pro tier — the previous MiMo-V2-Pro was API-only with no public weights. Weights live at huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro under the MIT license. The model card was last updated April 28, 2026, so deployment tooling, community quants, and reproductions are still landing day-by-day.

MiMo-V2.5-Pro is a 1.02-trillion parameter Mixture-of-Experts model that activates only ~42B parameters per token. The MiMo team — led by ex-DeepSeek researcher Luo Fuli — designed it around two ideas: a hybrid attention scheme that blends Sliding Window Attention (SWA) and Global Attention (GA) at a 6:1 ratio (~7× KV-cache reduction with a 128-token window), and 3 lightweight Multi-Token Prediction (MTP) modules that yield roughly 3× output speed on autoregressive workloads. The architecture has 70 layers (1 dense + 69 MoE), hidden size 6144, and ships natively in FP8 E4M3 mixed precision.

Two things matter for Clore.ai users. First, this is the first MiMo Pro release with public weights: previous Pro variants only existed as a hosted API and as the stealth-tested "Hunter Alpha" model on OpenRouter (March 2026 timeline). Second, the MIT license removes commercial restrictions outright — fine-tune, redistribute, run it as a paid endpoint, no caveats. Xiaomi's launch announcement claims V2.5-Pro beats DeepSeek V4 on agentic tasks, but that benchmark is vendor-published only — third-party reproduction has not landed yet, and you should not quote it externally without that caveat.

Key Specs

Property	Value
Total Parameters	1.02T (MoE)
Active Parameters	~42B per forward pass
Context Window	1,000,000 tokens (1M)
Precision	FP8 E4M3 mixed (native)
Architecture	Hybrid SWA + GA (6:1), 70 layers (1 dense + 69 MoE), hidden 6144
KV-Cache	Sliding window 128, ~7× reduction vs full GA
Speculative Decoding	3 lightweight MTP modules, ~3× output speed
License	MIT
Release Date	April 27, 2026
Organization	Xiaomi MiMo team (XiaomiMiMo on HuggingFace)
Primary Tooling	SGLang (first-class), vLLM

Why MiMo-V2.5-Pro?

• First open Pro-tier MiMo — predecessor MiMo-V2-Pro was API-only, this is the first time the Pro weights are public

• 1M-token context — handles entire codebases, long agent traces, or multi-document RAG without chunking

• Hybrid attention — SWA + GA at 6:1 cuts KV-cache ~7× vs pure global attention; long contexts stay tractable

• Native FP8 — no post-hoc quantization, weights ship in FP8 E4M3 directly from the vendor

• MTP speculative decoding — 3 built-in MTP modules give ~3× decode throughput out of the box

• MIT license — no commercial restrictions, no field-of-use limits

• 42B active — you pay 42B-dense inference cost despite the 1.02T headline number

• Lineage — lead researcher Luo Fuli was previously at DeepSeek, and the architectural choices show

---

Requirements

Still a 1T model. "42B active" sounds friendly, but the full 1.02T weights must live in VRAM (or be aggressively offloaded). Native FP8 weights need ~600GB+ VRAM before activation memory and KV cache. Plan for 8×H200 or larger for full-context FP8.

Component	Minimum (Quant + offload, future)	Recommended (FP8)	Full FP8, 1M ctx
GPU VRAM	~141GB (Q4 + RAM offload, when quants land)	8× H100 80GB (640GB)	8× H200 141GB (1,128GB)
RAM	256GB	512GB	512GB
Disk	700GB NVMe	1.5TB NVMe	2TB NVMe
CUDA	12.4+	12.6+	12.6+

Clore.ai pick: For full FP8 with breathing room on the 1M context, 8×H200 is the natural fit — see clore.ai/rent-h200.html. 8×H100 80GB also runs the FP8 checkpoint but you'll cap --context-length lower (typically 256K) to leave room for KV cache. For Blackwell-class hardware see clore.ai/rent-b200.html.

---

Option A — Ollama / GGUF (Quantized, community builds)

Heads-up: As of April 28, 2026 (one day after release) community GGUF quants for MiMo-V2.5-Pro are not yet published. Expect Q4_K_M / Q5_K_M / Q6_K builds to appear within 1–2 weeks at huggingface.co/models?search=mimo-v2.5-pro+gguf. Until then, FP8 via SGLang or vLLM is the supported path.

# Once a Q4_K_M build is available
docker exec ollama ollama pull mimo-v2.5-pro:q4_K_M
docker exec ollama ollama run mimo-v2.5-pro:q4_K_M

# Or with llama.cpp directly on a GGUF file (when published)
docker run --gpus all -it --rm -p 8080:8080 \
  -v $(pwd)/models:/models \
  ghcr.io/ggerganov/llama.cpp:server-cuda \
  -m /models/mimo-v2.5-pro-q4_k_m.gguf \
  --n-gpu-layers 99 --ctx-size 65536 \
  --port 8080 --host 0.0.0.0

---

Option B — vLLM (Production API, recommended)

vLLM supports MiMo-V2.5-Pro via --trust-remote-code (the hybrid attention + MTP modules ship as custom code in the repo). Use the vendor sampling defaults: temperature 1.0, top_p 0.95.

version: "3.8"
services:
  vllm:
    image: vllm/vllm-openai:latest
    ports:
      - "8000:8000"
    volumes:
      - hf_cache:/root/.cache/huggingface
    command: >
      --model XiaomiMiMo/MiMo-V2.5-Pro
      --tensor-parallel-size 8
      --quantization fp8
      --max-model-len 262144
      --gpu-memory-utilization 0.90
      --trust-remote-code
      --served-model-name mimo-v2.5-pro
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    shm_size: "16gb"

volumes:
  hf_cache:

# Test the API (vendor-recommended sampling)
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mimo-v2.5-pro",
    "messages": [
      {"role": "system", "content": "You are an autonomous coding agent."},
      {"role": "user", "content": "Walk through this 30K-line monorepo and propose a migration plan from Express 4 to Fastify 5."}
    ],
    "max_tokens": 8192,
    "temperature": 1.0,
    "top_p": 0.95
  }'

On 8×H100 80GB, cap --max-model-len at 262144 (256K) to leave headroom for activations + KV cache. On 8×H200 141GB you can comfortably push to 524288 or higher; 1,048,576 (full 1M) is feasible but expect long prefill times — test before relying on it.

---

Option C — SGLang (recommended for max throughput)

SGLang is the first-class serving target in the MiMo-V2.5-Pro model card. The vendor publishes the launch command with SGLANG_ENABLE_SPEC_V2=1 to activate the new MTP-aware speculative decoding path, which is where the ~3× decode speedup actually materializes.

docker pull lmsysorg/sglang:latest

# Verbatim from the HF model card
SGLANG_ENABLE_SPEC_V2=1 python3 -m sglang.launch_server \
    --model-path XiaomiMiMo/MiMo-V2.5-Pro \
    --trust-remote-code \
    --quantization fp8 \
    --context-length 1048576 \
    --host 0.0.0.0 --port 9001

For a multi-GPU TP setup on 8×H200, add --tp-size 8 and --mem-fraction-static 0.88. Confirm with nvidia-smi that all 8 cards are populated before sending real traffic — the 1M context is unforgiving if one rank is starved.

---

Clore.ai GPU Recommendations

Setup	VRAM	Expected Performance	Clore.ai Cost
4× H100 80GB	320GB	FP8 with heavy offload, max ctx ~64K, ~10–15 tok/s	~$25–35/day
8× H100 80GB	640GB	FP8 full, max ctx ~256K, ~30–45 tok/s	~$45–60/day
8× H200 141GB	1,128GB	FP8 full, max ctx 1M, ~60+ tok/s with MTP	~$80–110/day
8× B200	1,536GB	FP8 full, max ctx 1M, fastest available	marketplace pricing

Best value: 8× H200 141GB on the FP8 checkpoint with SGLANG_ENABLE_SPEC_V2=1. You get the full 1M context window, MTP speculative decoding, and enough KV-cache headroom for real agent loops. See clore.ai/rent-h200.html for live availability.

---

Use Cases

• Long-horizon agents — MiMo team explicitly tunes for sustained tool-calling. The 1M context plus MTP speedup means thousands of tool turns without chunking gymnastics.

• Whole-codebase analysis — drop a 500K-token monorepo into context for refactor planning, dependency audits, or migration design

• Long-document RAG — entire books, multi-year customer transcripts, or year-long chat histories fit in one prompt

• Coding — vendor-claimed HumanEval+ 75.6% and the agentic posture make it a candidate for autonomous SWE workloads (pair with SWE-agent / OpenHands)

• Research scratchpad — 1M context tolerates the kind of "paste the whole paper, paste the prior work, ask for synthesis" usage that smaller models truncate

---

Benchmarks

Vendor-claimed — no third-party reproduction yet. All numbers below come from Xiaomi's April 27, 2026 announcement and the HuggingFace model card. The model is two days old at time of writing — independent reproductions on agentic and long-context benchmarks are still pending. The "beats DeepSeek V4 on agentic tasks" claim in particular is from Xiaomi's own write-up; treat it as marketing until reproduced.

Benchmark	MiMo-V2.5-Pro (vendor)	Notes
GSM8K	99.6%	Math word problems
HumanEval+	75.6%	Coding (extended)
MMLU	89.4%	General knowledge
GraphWalks (1M ctx) BFS	0.37	Long-context graph traversal
GraphWalks (1M ctx) Parents	0.62	Long-context graph traversal
Agentic tasks vs DeepSeek V4	"outperforms" (vendor)	Unverified — third-party reproduction pending

---

Troubleshooting

Issue	Solution
OutOfMemoryError on load	Native FP8 still needs ~600GB+ VRAM. Use 8× H200 or drop --context-length to 65536 on 8× H100.
Slow HuggingFace download	huggingface-cli download XiaomiMiMo/MiMo-V2.5-Pro --local-dir ./weights --resume-download. Expect ~600GB FP8.
--trust-remote-code rejected	Hybrid attention and MTP ship as custom code in the repo. The flag is mandatory for both vLLM and SGLang.
MTP speedup not appearing in SGLang	Confirm SGLANG_ENABLE_SPEC_V2=1 is exported in the same shell as python3 -m sglang.launch_server. The default path does not activate MTP.
Reasoning trace flat / low quality	Use temperature=1.0 and top_p=0.95. Lower temps degrade MiMo's reasoning behavior.
1M context OOMs on 8× H100	8× H100 80GB cannot hold KV cache for 1M tokens. Cap at 256K or move to 8× H200.
Prefill takes minutes	Expected at 1M context. Use --enable-chunked-prefill (vLLM) or batch shorter requests for interactive workloads.
GGUF / Ollama pull fails	Community quants are not published as of April 28, 2026. Wait 1–2 weeks or use FP8 directly.

---

Next Steps

• Predecessor / sibling: MiMo-V2-Flash — 309B MoE, 15B active, 32K ctx, faster but smaller

• Vendor's claimed rival: DeepSeek V4 — 1M ctx, multimodal, ~1T params (the model Xiaomi says they beat on agentic tasks)

• Open-weight coding rival: GLM-5.1 — 744B MoE, 40B active, MIT, currently #1 on SWE-Bench Pro

• Clore.ai H200 rentals: clore.ai/rent-h200.html — best fit for full FP8 1T MoE at 1M context

• Clore.ai marketplace: clore.ai/marketplace

Links

• MiMo-V2.5-Pro on HuggingFace

• Xiaomi MiMo HuggingFace org

• SGLang repo

• vLLM docs