MiMo v2.5 Pro vs MiMo v2.6 Pro
AutomationBench-AA AutomationBench-AA: Agents orchestrate SaaS workflows through APIs; objectives receive credit only when the task has no guardrail violation
MiMo v2.5 Pro13.7
MiMo v2.6 Pro58.6
GDPval-AA (work tasks) 2.1 GDPval-AA (work tasks) 2.1: Real work tasks from many professions (reports, spreadsheets, analyses), graded against expert work
MiMo v2.5 Pro31
MiMo v2.6 Pro59.2
AA Analyst Agent AA Analyst Agent: Data-analyst tasks: the model explores data with tools and reports findings
MiMo v2.5 Pro20
MiMo v2.6 Pro37.5
Terminal-Bench 4.0 Terminal-Bench 4.0 · source setting 2: Agents use a terminal to complete executable coding, system administration and research tasks in the 4.0 suite
MiMo v2.5 Pro0
MiMo v2.6 Pro34.8
AA-Briefcase 1.1 · Elo AA-Briefcase 1.1 · Elo: Agents produce professional deliverables; blind pairwise comparisons produce the source Elo rating
MiMo v2.5 Pro882.8
MiMo v2.6 Pro1,514.7
AA-LCR (long context) 1.1 AA-LCR (long context) 1.1: Questions that need reading and connecting facts across very long documents (around 100K tokens)
MiMo v2.5 Pro79.7
MiMo v2.6 Pro86.3
τ³-Banking · AA τ³-Banking · AA: Agents retrieve banking knowledge and execute tools to resolve customer-support workflows
MiMo v2.5 Pro9.9
MiMo v2.6 Pro—
ITBench SRE ITBench SRE: IT incident tasks: diagnose and fix broken systems like a site-reliability engineer
MiMo v2.5 Pro38.2
MiMo v2.6 Pro—
APEX-Agents APEX-Agents: Agents complete long-horizon professional-services tasks across realistic applications
MiMo v2.5 Pro2.4
MiMo v2.6 Pro—
Terminal-Bench Hard (version unstated) Terminal-Bench Hard (version unstated): Agents solve difficult software, system-administration and data-processing tasks in a terminal
MiMo v2.5 Pro43.2
MiMo v2.6 Pro—
Terminal-Bench Science 0.1 Terminal-Bench Science 0.1: Agents complete expert-authored research workflows across five scientific domains in a terminal
MiMo v2.5 Pro—
MiMo v2.6 Pro5.7
FrontierCode 1.1 FrontierCode 1.1: Agents implement changes in real repositories, graded against maintainer-defined requirements
MiMo v2.5 Pro—
MiMo v2.6 Pro45.4
| Benchmark | MiMo v2.5 Pro | MiMo v2.6 Pro |
|---|---|---|
| AutomationBench-AA | 13.7 | 58.6 |
| GDPval-AA (work tasks) 2.1 | 31 | 59.2 |
| AA Analyst Agent | 20 | 37.5 |
| Terminal-Bench 4.0 | 0 | 34.8 |
| AA-Briefcase 1.1 · Elo | 882.8 | 1,514.7 |
| AA-LCR (long context) 1.1 | 79.7 | 86.3 |
| τ³-Banking · AA | 9.9 | — |
| ITBench SRE | 38.2 | — |
| APEX-Agents | 2.4 | — |
| Terminal-Bench Hard (version unstated) | 43.2 | — |
| Terminal-Bench Science 0.1 | — | 5.7 |
| FrontierCode 1.1 | — | 45.4 |
Specs
| Spec | MiMo v2.5 Pro | MiMo v2.6 Pro |
|---|---|---|
| Input, per 1M | $0.43 | $0.43 |
| Output, per 1M | $0.87 | $0.87 |
| Cost per task? | $0.054 | $0.13 |
| Speed, tok/s | 46 | 54 |
| Context | 1M | 1M |
| Weights | Open | Open |
| Released | Apr 2026 | Sep 2026 |
| Memory (4-bit) | 614 GB | 615 GB |
| Licence | MIT | MIT |
| Commercial use | Allowed | Allowed |
| See images | — | ✓ |
| Read PDFs | — | — |
| Hear audio | — | ✓ |
| Watch video | — | ✓ |
| Call tools | ✓ | ✓ |
| Force a tool call | Partial | Partial |
| Return strict JSON | ✓ | ✓ |
| Search the web | ? | ? |
| Run code | ? | ? |
| Use a computer | ? | ? |
What breaks when you switch
MiMo v2.5 Pro → MiMo v2.6 Pro
Memory (4-bit): 614 GB → 615 GB
Pass back reasoning_content›
In deep-thinking multi-turn tool conversations, return the full reasoning_content in later rounds with tool calls or receive HTTP 400.
Forced recommended defaults›
Deep thinking overrides supplied temperature and top_p with 1.0 and 0.95.
Avoid answer truncation›
max_completion_tokens includes thinking and the final answer; reserve enough tokens to avoid truncating the answer.
MOPD checkpoints address tool-call repetition›
Use the MOPD checkpoints for the tool-call repetition fix; the API has served the updated models since September 25 at 06:00 UTC+8 under unchanged names.
MiMo v2.6 Pro → MiMo v2.5 Pro