MiMo-V2.6-Pro API
Xiaomi's flagship MiMo-V2.6 model, a 1.02T-parameter sparse MoE trained with large-scale mixed reinforcement learning for demanding agentic work such as long-horizon coding, multi-tool agents and cybersecurity tasks.
xiaomi/mimo-v2.6-pro
Top up $20 and we add 50%. That gives you about 23.0M input tokens extra on this model, on us.
Create an accountPricing
USD per 1M tokens| Input | $0.44 |
| Output | $0.87 |
| Cached input | $0.0099% below fresh input |
Billed per token with no minimum and no monthly fee. Price your workload
Capabilities and limits
- Context window
- 1M tokens
- Max output
- 128K tokens
- Modality
- multimodal
- Tool calling
- Supported
- Vision
- Supported
- Reasoning
- Optional
- Prompt caching
- Supported (automatic)
Measured performance
7d- Time to first token (p50)
- 400 ms
- Time to first token (p95)
- 400 ms
- Throughput (avg)
- 102.7 tok/s
- Throughput (p50)
- 150.0 tok/s
Measured on real Run BiOS traffic over the trailing 7d. Your figures will vary with prompt shape and region.
curl https://api.runbios.ai/v1/chat/completions \
-H "Authorization: Bearer $BIOS_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "xiaomi/mimo-v2.6-pro",
"messages": [{ "role": "user", "content": "Hello" }]
}'Existing OpenAI SDK code works by changing the base URL and the key. API overview
Model card
Curated by Run BiOS · benchmark figures, where present, are the card’s own, not our measurements
MiMo-V2.6-Pro is the flagship model of the MiMo-V2.6 series from Xiaomi's MiMo team, a sparse Mixture-of-Experts (MoE) model with 1.02T total and 42B activated parameters, a 1-million-token context window and native understanding of text, images, video and audio. Xiaomi built the V2.6 series to scale reinforcement learning toward self-improvement: RL compute, environment diversity and grader compute are scaled together, so the model keeps expanding its capability frontier through exploration and feedback. Xiaomi positions the Pro model for demanding agentic work, including long-horizon coding, multi-tool agents and cybersecurity tasks. It is released under the MIT License.
Key features reported by Xiaomi:
- Native omnimodal and long horizon. Text, image, video and audio in one model, with 1M tokens of context for long repositories, tool traces and multi-session agent runs.
- You Only RL Once. One mixed reinforcement learning run across coding, general agents, visual tasks and cybersecurity rather than separate per-domain runs. Tasks and multiple agent harnesses are mixed in the same batch, so capabilities reinforce each other and strategies transfer to harnesses never seen in training.
- Scaling RL compute. Fully asynchronous Group Relative Policy Optimization (GRPO) on very large batches: 1,568 prompts with 16 rollouts each per step, billions of tokens per update.
- Groupwise agentic grading. A binary pass or fail cannot rank passing solutions, so Xiaomi scales the reward signal itself. An agentic grader compares rollouts within each group. Groupwise Reward Synthesis (GRS) builds task-specific rubrics offline from contrasting rollouts and fuses rubric quality with test outcomes. Groupwise Advantage Redistribution (GAR) ranks passing trajectories online and moves advantage toward higher-quality solutions. Judged against the policy's own samples, this closes a self-improvement loop and steers the model toward shorter paths and fewer tokens per task.
- Aligned RL. Training starts from self-correction: the model reflects on its own misaligned turns and rewrites them into grounded next steps. Throughout RL, environment hardening, adversarial screening and cross-checks between verifiers guard the loop against reward hacking.
- Multi-Prefix Multi-Teacher On-Policy Distillation (MOPD2). After mixed RL, MOPD2 combines autonomous student rollouts with prefix-conditioned single-turn rollouts that reuse histories from teacher trajectories and supervised demonstrations. Decision points are trained without regenerating the preceding turns, which extends the model's capabilities to tasks that are hard to verify.
Xiaomi has also published a MOPD upgrade of the Pro checkpoint. It fuses several domain-specialized teachers into one model, extending to domains where reliable training-time verification is hard, such as long-horizon game development, scientific research and embodied intelligence. It also targets tool-call repetition, a failure mode Xiaomi identified after the V2.6 release in which the model keeps issuing the same or highly similar tool calls, appearing busy while making no progress.
Benchmarks
Reproduced from Xiaomi's official MiMo-V2.6 model card. All numbers are as published. "n/a" means the source table does not report a score.
| Benchmark | MiMo-V2.6 Pro | MiMo-V2.6 Flash | MiMo-V2.5 Pro | Claude Opus 5 | GPT-5.6 Sol | Claude Fable 5 |
|---|---|---|---|---|---|---|
| Code Agent | ||||||
| DeepSWE v1.1 | 71.9 | 67.9 | 19.0 | 74.0 | 73.0 | 70.0 |
| ProgramBench | 26.5 | 26.0 | 12.5 | 37.0 | 25.0 | 33.0 |
| MiMo Code Bench | 63.2 | 61.2 | 40.4 | 68.6 | 59.3 | n/a |
| General Agent | ||||||
| AutomationBench v1.0.6 | 53.1 | 52.3 | 16.0 | 50.3 | 45.8 | 46.2 |
| Toolathlon-Verified | 76.9 | 73.6 | 49.1 | 80.6 | 74.9 | 77.9 |
| GDPval-AA 2.1 | 1673 | n/a | 1107 | 1708 | 1588 | 1595 |
| Agents' Last Exam | 31.6 | 27.6 | 13.2 | 31.6 | 30.8 | 25.7 |
| Terminal Bench 4.0 | 34.9 | 28.8 | 1.5 | 49.0 | 39.9 | 42.4 |
| Terminal Bench 2.1 | 89.9 | 87.6 | 65.2 | 89.1 | 88.8 | 84.3 |
| OSWorld-Verified | 82.0 | 80.8 | n/a | 83.4 | 83.0 | 86.0 |
| JobBench | 62.0 | 61.2 | 25.0 | 65.7 | 45.4 | 57.4 |
| Cybersecurity | ||||||
| CyberGym | 94.0 | 95.1 | 40.0 | n/a | n/a | n/a |
| MiMo Cyber Bench | 80.2 | 77.2 | 0.0 | n/a | n/a | n/a |
| ExploitGym | 17.8 | 6.0 | 0.2 | 22.1 | 30.3 | 28.4 |
| ExploitBench | 47.9 | 25.3 | 16.6 | 70.0 | 78.5 | 78.0 |
| SEC Bench Pro | 66.3 | 47.5 | 17.7 | n/a | 79.1 | n/a |
| Visual Agent | ||||||
| MiMo VisualCoding | 72.3 | 71.5 | n/a | 70.0 | 73.4 | 69.1 |
Against MiMo-V2.5-Pro, the V2.6 Pro model scores higher on every benchmark where both are reported, with the largest gains on MiMo Cyber Bench (0.0 to 80.2), CyberGym (40.0 to 94.0) and DeepSWE v1.1 (19.0 to 71.9). In this table it has the top score on AutomationBench v1.0.6, Terminal Bench 2.1 and MiMo Cyber Bench and it ties Claude Opus 5 on Agents' Last Exam. Claude Opus 5 leads DeepSWE v1.1, ProgramBench, MiMo Code Bench, Toolathlon-Verified, GDPval-AA 2.1, Terminal Bench 4.0 and JobBench. GPT-5.6 Sol leads ExploitGym, ExploitBench, SEC Bench Pro and MiMo VisualCoding. Claude Fable 5 leads OSWorld-Verified and MiMo-V2.6-Flash leads CyberGym.
Architecture
All architecture figures are from Xiaomi's official model card.
Model summary
- Architecture: sparse MoE, 1.02T total and 42B activated parameters
- Context length: 1M tokens
- Modalities: text, image, video and audio
- Vision encoder: 681M-parameter MiMo ViT (28 layers: 24 sliding-window and 4 full attention)
- Audio encoder: 308M-parameter AudioTokenizer plus a 127M-parameter audio patch encoder
- Multi-Token Prediction (MTP): 5-layer speculative decoder
LLM backbone
| Component | MiMo-V2.6-Pro |
|---|---|
| Layers (Total / SWA / GA) | 70 / 60 / 10 |
| Hidden Size | 6144 |
| SWA Heads (Q/KV) | 128 / 8 |
| GA Heads (Q/KV) | 128 / 8 |
| Head Dimensions (QK / V) | 192 / 128 |
| Sliding Window Size | 128 |
| Routed Experts (Total / Activated) | 384 / 8 |
| Max Context Length | 1M |
| MTP / Speculative Decoder | 5 SWA layers, window 1024 |
The first Transformer block uses global attention with a dense FFN. The remaining blocks interleave local sliding-window attention (SWA) and global attention (GA), and both use sparse MoE FFNs without shared experts.
Vision encoder (MiMo ViT)
| Configuration | Value |
|---|---|
| Layers (Total / SWA / GA) | 28 / 24 / 4 |
| Hidden Size | 1280 |
| Attention Heads (Q / KV) | 32 / 8 |
| Head Dimension | 64 |
| Patch Size (T x H x W) | 2 x 16 x 16 |
| Sliding Window (Left / Right) | 64 / 64 |
| Spatial Merge Size | 2 x 2 |
| Parameters | 681M |
Audio encoders. The AudioTokenizer encoder has 24 layers (12 SWA and 12 GA), hidden size 1024, 20 RVQ codebooks and 308M parameters. The audio patch encoder has 6 layers and 127M parameters and groups four frames per patch (25 Hz to 6.25 Hz).
Speculative decoder. A 5-layer SWA MTP drafter predicts 7 subsequent tokens per forward pass for parallel verification.
Reasoning and Sampling
MiMo-V2.6-Pro is a reasoning model whose thinking can be switched off. On Run BiOS reasoning is on by default and can be turned off per request with reasoning_effort: "none". Xiaomi's recommended sampling parameters are temperature = 1.0 and top_p = 0.95.
Usage Restrictions
Xiaomi's model card asks users not to use the model or its derivatives:
- for any unlawful, fraudulent or malicious purpose
- to infringe the intellectual property, privacy or other rights of others
- to harm, harass or unlawfully discriminate against people, including minors and vulnerable people
- for any military purpose
- for malicious cyber activity such as malware, ransomware, phishing, credential theft or unauthorized access
- to access or process personal data without a lawful basis or automate other platforms in violation of their policies
- to execute high-risk actions autonomously without appropriate human oversight
Xiaomi notes that this notice is an informational reminder and does not change the MIT License.
On Run BiOS
MiMo-V2.6-Pro is available on Run BiOS as the serverless model xiaomi/mimo-v2.6-pro, with a 1,048,576-token context window, up to 131,072 output tokens, image input, function calling, structured outputs, optional reasoning and automatic prompt caching. The Run BiOS API accepts text and image input for this model. Audio input is not offered.