MiMo-V2.5-Pro: 1.02T parameters, MIT-licensed agent powerhouse

From V2-Pro to V2.5-Pro: the long-context breakthrough

XiaoMi’s MiMo family has rapidly positioned itself among the top open-weight models. MiMo-V2.5-Pro is the latest iteration — a 1.02 trillion-parameter Mixture of Experts model with 42B active parameters that significantly extends what was possible in MiMo-V2-Pro. The headline capability is not just scale, but long-context persistence at 1M tokens combined with agentic reasoning trained via multi-teacher on-policy distillation.

Where MiMo-V2-Pro proved the architecture could handle long sequences, V2.5-Pro proves it can retain and apply information across a full million tokens while maintaining agentic reasoning. On long-context graph navigation benchmarks, V2.5-Pro scores 0.62 on the Parents task at 1M tokens (vs. V2-Pro collapsing to 0 at that length), a qualitative leap in context retention.

---

Architecture: MoE with Hybrid Attention and Multi-Token Prediction

MiMo-V2.5-Pro uses an 70-layer architecture (1 dense + 69 MoE) with a carefully structured attention system and novel prediction heads.

Hybrid Attention (SWA + GA, 6:1 ratio).
60 layers use Sliding Window Attention (SWA) with a 128-token window, handling local context efficiently. The remaining 10 layers use Global Attention (GA) to capture long-range dependencies. The 6:1 ratio was found optimal during pre-training — enough global access to prevent information loss across 1M tokens, while the sliding window keeps per-layer compute manageable.

Mixture of Experts (384 experts, 8 routed per token).
The model has 384 total experts with 8 active per token, yielding 42B active parameters out of 1.02T total. The routing network distributes tokens across experts dynamically, with each expert specialising in different linguistic or reasoning patterns discovered during the 27T-token pre-training phase.

Multi-Token Prediction (3 layers).
Three MTP heads are attached to intermediate layers, allowing the model to predict the next 3 tokens simultaneously during training. This accelerates convergence and improves the quality of the learned representations without adding inference overhead, as MTP is only used during pre-training.

FP8 mixed-precision training.
MiMo-V2.5-Pro was trained in FP8 mixed precision, a first at this model scale. FP8 reduces memory bandwidth requirements and increases throughput during training without noticeable accuracy loss — the 27T-token pre-training phase ran at 32K sequence length using FP8, with a switch to BF16 for the final post-training stages.

---

Benchmark results

Base model evaluation (post pre-training, before fine-tuning)

Benchmark	MiMo-V2.5-Pro Base	MiMo-V2-Pro Base
BBH	88.4	85.1
MMLU	89.4	87.3
MMLU-Pro	68.5	64.2
MATH	86.2	83.8
HumanEval+	75.6	71.3
SWE-Bench (AgentLess)	35.7	31.2
C-Eval	91.5	89.8

The base model already shows a strong gap over V2-Pro across all categories, reflecting the benefit of 27T pre-training tokens and FP8 training efficiency.

Post-training evaluation (SFT + agentic RL + MOPD)

Benchmark	MiMo-V2.5-Pro	MiMo-V2-Pro
SWE-Bench Pro	57.2%	—
SWE-Bench Verified	78.9%	74.1%
TerminalBench 2	68.4%	59.3%
GPQA-Diamond	66.7%	63.5%
GSM8K	99.6%	98.7%

The post-training gains are substantial — particularly on SWE-Bench Pro (+~4 points) and TerminalBench 2 (+9.1 points). The Multi-Teacher On-Policy Distillation (MOPD) plays a key role here: multiple teacher models guide the student through on-policy trajectories, teaching it to reason more systematically in agentic scenarios.

---

Long-context benchmark

V2.5-Pro’s defining achievement is preserving information at extreme context lengths.

Benchmark	MiMo-V2.5-Pro (1M)	MiMo-V2.5 (1M)	MiMo-V2-Pro (collapse)
GraphWalks BFS	0.37	0.31	0.00
GraphWalks Parents	0.62	0.48	0.00

At 1M tokens, V2-Pro completely collapses (0.00 on both tasks), while V2.5 retains meaningful ability and V2.5-Pro achieves the strongest scores. This is not a marginal improvement — it is the difference between a model that can reason across a 2000-page document and one that cannot.

---

What makes it different: long-context persistence + agentic capability

Two capabilities set MiMo-V2.5-Pro apart from competitors:

1. Long-context that actually works.
Hybrid Attention with the 6:1 SWA/GA ratio and MTP training together enable information to travel across 1M tokens. Most models lose context coherence well before 256K. V2.5-Pro was explicitly evaluated and validated on graph navigation tasks spanning millions of tokens — a task that requires remembering relationships across an arbitrarily long sequence.

2. Agentic reasoning via MOPD.
Post-training with Multi-Teacher On-Policy Distillation gives V2.5-Pro the ability to plan, execute multi-step operations (code, search, browse), and recover from errors autonomously. The agentic RL loop taught the model to maintain strategy across hundreds of action steps, which directly translates to SWE-Bench Pro scores of 57.2%.

---

Comparison with MiMo-V2.5 and MiMo-V2-Pro

Dimension	MiMo-V2-Pro	MiMo-V2.5	MiMo-V2.5-Pro
Total params	1.02T	1.02T	1.02T
Active params	42B	42B	42B
Architecture	MoE + Hybrid Attn	MoE + Hybrid Attn	MoE + Hybrid Attn + MTP
Layers	69	69	70 (1 dense + 69 MoE)
Experts	384 / 8	384 / 8	384 / 8
Context	128K	512K	1M
Pre-training tokens	14T	20T	27T
Pre-training precision	FP16	BF16	FP8 → BF16
Post-training	SFT	SFT + RL	SFT + Agentic RL + MOPD
SWE-Bench Pro	—	—	57.2%
GraphWalks Parents @ 1M	0.00	0.00	0.62
License	MIT	MIT	MIT

V2.5-Pro is the only model in the family with MTP training, the full 1M context validation, and the MOPD post-training pipeline.

---

Deployment

MiMo-V2.5-Pro can be deployed via two primary inference frameworks:

• SGLang (latest) with EAGLE speculative decoding for accelerated inference
• vLLM (latest) for high-throughput serving

An API is available at platform.xiaomimimo.com.

For local deployment with EAGLE (speculative decoding), SGLang provides significant speedups by using a smaller draft model to propose tokens, which the V2.5-Pro verifier accepts or rejects — typically achieving 2-3× throughput improvement without accuracy loss.

---

License and access

MiMo-V2.5-Pro is released under the MIT license — full commercial use, modification, and redistribution allowed with minimal restrictions.

The model is available at: XiaoMi/MiMo-V2.5-Pro

---

Citation

@misc{xiaomicom2026.mimov25pro,
  title={MiMo V2.5 Pro: Efficient Long-Context Language Model with Multi-Teacher On-Policy Distillation},
  author={XiaoMi},
  year={2026}
}