MiMo-V2-Flash is a Mixture-of-Experts (MoE) language model with 309B total parameters and 15B active parameters. Designed for high-speed reasoning and agentic workflows, it utilizes a novel hybrid attention architecture and Multi-Token Prediction (MTP) to achieve state-of-the-art performance while significantly reducing inference costs.
1. Introduction
MiMo-V2-Flash creates a new balance between long-context modeling capability and inference efficiency. Key features include:
Hybrid Attention Architecture: Interleaves Sliding Window Attention (SWA) and Global Attention (GA) with a 5:1 ratio and an aggressive 128-token window. This reduces KV-cache storage by nearly 6x while maintaining long-context performance via learnable attention sink bias.
Multi-Token Prediction (MTP): Equipped with a lightweight MTP module (0.33B params/block) using dense FFNs. This triples output speed during inference and will be good to accelerates rollout in RL training.
Efficient Pre-Training: Trained on 27T tokens using FP8 mixed precision and native 32k seq length. The context window supports up to 256k length.
Agentic Capabilities: Post-training utilizes Multi-Teacher On-Policy Distillation (MOPD) and large-scale agentic RL, achieving superior performance on SWE-Bench and complex reasoning tasks.
[!IMPORTANT]
We also open-source the 3-layer MTP weights to foster community research.
3. Evaluation Results
Base Model Evaluation
MiMo-V2-Flash-Base demonstrates strong performance across standard benchmarks, surpassing models with significantly larger parameter counts.
Category
Benchmark
Setting/Length
MiMo-V2-Flash Base
Kimi-K2 Base
DeepSeek-V3.1 Base
DeepSeek-V3.2 Exp Base
Params
#Activated / #Total
-
15B / 309B
32B / 1043B
37B / 671B
37B / 671B
General
BBH
3-shot
88.5
88.7
88.2
88.7
MMLU
5-shot
86.7
87.8
87.4
87.8
MMLU-Redux
5-shot
90.6
90.2
90.0
90.4
MMLU-Pro
5-shot
73.2
69.2
58.8
62.1
DROP
3-shot
84.7
83.6
86.3
86.6
ARC-Challenge
25-shot
95.9
96.2
95.6
95.5
HellaSwag
10-shot
88.5
94.6
89.2
89.4
WinoGrande
5-shot
83.8
85.3
85.9
85.6
TriviaQA
5-shot
80.3
85.1
83.5
83.9
GPQA-Diamond
5-shot
55.1
48.1
51.0
52.0
SuperGPQA
5-shot
41.1
44.7
42.3
43.6
SimpleQA
5-shot
20.6
35.3
26.3
27.0
Math
GSM8K
8-shot
92.3
92.1
91.4
91.1
MATH
4-shot
71.0
70.2
62.6
62.5
AIME 24&25
2-shot
35.3
31.6
21.6
24.8
Code
HumanEval+
1-shot
70.7
84.8
64.6
67.7
MBPP+
3-shot
71.4
73.8
72.2
69.8
CRUXEval-I
1-shot
67.5
74.0
62.1
63.9
CRUXEval-O
1-shot
79.1
83.5
76.4
74.9
MultiPL-E HumanEval
0-shot
59.5
60.5
45.9
45.7
MultiPL-E MBPP
0-shot
56.7
58.8
52.5
50.6
BigCodeBench
0-shot
70.1
61.7
63.0
62.9
LiveCodeBench v6
1-shot
30.8
26.3
24.8
24.9
SWE-Bench (AgentLess)
3-shot
30.8
28.2
24.8
9.4*
Chinese
C-Eval
5-shot
87.9
92.5
90.0
91.0
CMMLU
5-shot
87.4
90.9
88.8
88.9
C-SimpleQA
5-shot
61.5
77.6
70.9
68.0
Multilingual
GlobalMMLU
5-shot
76.6
80.7
81.9
82.0
INCLUDE
5-shot
71.4
75.3
77.2
77.2
Long Context
NIAH-Multi
32K
99.3
99.8
99.7
85.6*
64K
99.9
100.0
98.6
85.9*
128K
98.6
99.5
97.2
94.3*
256K
96.7
-
-
-
GSM-Infinite Hard
16K
37.7
34.6
41.5
50.4
32K
33.7
26.1
38.8
45.2
64K
31.5
16.0
34.7
32.6
128K
29.0
8.8
28.7
25.7
* indicates the model may fail to follow the prompt or format.
Post-training Model Evaluation
Following our Post-Training Paradigm with MOPD and Agentic RL, the model achieves SOTA reasoning and agentic performance.
Benchmark
MiMo-V2 Flash
Kimi-K2 Thinking
DeepSeek-V3.2 Thinking
Gemini-3.0 Pro
Claude Sonnet 4.5
GPT-5 High
Reasoning
MMLU-Pro
84.9
84.6
85.0
90.1
88.2
87.5
GPQA-Diamond
83.7
84.5
82.4
91.9
83.4
85.7
HLE (no tools)
22.1
23.9
25.1
37.5
13.7
26.3
AIME 2025
94.1
94.5
93.1
95.0
87.0
94.6
HMMT Feb. 2025
84.4
89.4
92.5
97.5
79.2
88.3
LiveCodeBench-v6
80.6
83.1
83.3
90.7
64.0
84.5
General Writing
Arena-Hard (Hard Prompt)
54.1
71.9
53.4
72.6
63.3
71.9
Arena-Hard (Creative Writing)
86.2
80.1
88.8
93.6
76.7
92.2
Long Context
LongBench V2
60.6
45.1
58.4
65.6
61.8
-
MRCR
45.7
44.2
55.5
89.7
55.4
-
Code Agent
SWE-Bench Verified
73.4
71.3
73.1
76.2
77.2
74.9
SWE-Bench Multilingual
71.7
61.1
70.2
-
68.0
55.3
Terminal-Bench Hard
30.5
30.6
35.4
39.0
33.3
30.5
Terminal-Bench 2.0
38.5
35.7
46.4
54.2
42.8
35.2
General Agent
BrowseComp
45.4
-
51.4
-
24.1
54.9
BrowseComp (w/ Context Manage)
58.3
60.2
67.6
59.2
-
-
\(\tau^2\)-Bench
80.3
74.3
80.3
85.4
84.7
80.2
4. Model Architecture
Hybrid Sliding Window Attention
MiMo-V2-Flash addresses the quadratic complexity of long contexts by interleaving Local Sliding Window Attention (SWA) and Global Attention (GA).
Configuration: Stacks of \(M=8\) hybrid blocks. Each block contains \(N=5\) SWA layers followed by 1 GA layer.
Efficiency: SWA layers use a window size of 128 tokens, reducing KV cache significantly.
Sink Bias: Learnable attention sink bias is applied to maintain performance despite the aggressive window size.
Lightweight Multi-Token Prediction (MTP)
Unlike traditional speculative decoding, our MTP module is natively integrated for training and inference.
Structure: Uses a dense FFN (instead of MoE) and SWA (instead of GA) to keep the parameter count low (0.33B per block).
Performance: Facilitates self-speculative decoding, tripling generation speed and mitigating GPU idleness during small-batch RL training.
5. Post-Training Technical Highlights
MiMo-V2-Flash leverages a post-training pipeline designed to maximize reasoning and agentic capabilities through innovative distillation and reinforcement learning strategies.
5.1 Multi-Teacher On-Policy Distillation (MOPD)
We introduce Multi-Teacher On-Policy Distillation (MOPD), a new paradigm that formulates knowledge distillation as a reinforcement learning process.
Dense Token-Level Guidance: Unlike methods relying on sparse sequence-level feedback, MOPD utilizes domain-specific expert models (teachers) to provide supervision at every token position.
On-Policy Optimization: The student model learns from its own generated responses rather than a fixed dataset. This eliminates exposure bias and ensures smaller, more stable gradient updates.
Inherent Reward Robustness: Rewards are derived from the distribution divergence between student and teacher, making the process naturally resistant to reward hacking.
5.2 Scaling Agentic RL
We significantly scale up the agentic training environments to improve intelligence and generalization.
Massive Code Agent Environments: We utilize real-world GitHub issues to create over 100,000 verifiable tasks. Our automated pipeline maintains a Kubernetes cluster capable of running over 10,000 concurrent pods with a 70% environment setup success rate.
Multimodal Verifier for WebDev: For web development tasks, we employ a vision-based verifier that evaluates code execution via recorded videos rather than static screenshots. This reduces visual hallucination and ensures functional correctness.
Cross-Domain Generalization: Our experiments show that large-scale RL training on code agents effectively generalizes to other domains, boosting performance in Math and General Agent tasks.
5.3 Advanced RL Infrastructure
To support high-throughput RL training for large-scale MoE models, we implemented several infrastructure optimizations on top of SGLang and Megatron-LM.
Rollout Routing Replay (R3): Addresses numerical precision inconsistencies in MoE routing between inference and training. R3 reuses the exact routed experts from rollout during the training pass, ensuring consistency with negligible overhead.
Request-Level Prefix Cache: In multi-turn agent training, this cache stores KV states and routed experts from prior turns. It avoids re-computation and ensures sampling consistency across turns.
Fine-Grained Data Scheduler: We extend the rollout engine to schedule fine-grained sequences instead of micro-batches. Combined with partial rollout, this significantly reduces GPU idleness caused by long-tail stragglers.
Toolbox & Tool Manager: A two-layer design using Ray actor pools to handle resource contention. It eliminates cold-start delays for tool execution and isolates task logic from system policies.
6. Inference & Deployment
MiMo-V2-Flash supports FP8 mixed precision inference. We recommend using SGLang for optimal performance.
temperature=0.3 for agentic taks (e.g., vibe-coding, tool-use)
3. Tool-use practice
[!IMPORTANT]
In the thinking mode with multi-turn tool calls, the model returns a reasoning_content field alongside tool_calls. To continue the conversation, the user must persist all history reasoning_content in the messages array of each subsequent request.
7. Citation
If you find our work helpful, please cite our technical report:
