Xiaomi
MiMo-V2.6-Pro
Xiaomi's new open-weight flagship: a 1.02T-parameter sparse MoE with MIT weights, 1M context and API pricing of $0.435/$0.87 per million tokens. Four of its seven tracked scores are independent: Humanity's Last Exam 49.35, AA-AnalystAgent 37.5 pass^5 (both Artificial Analysis), vals.ai's Terminal-Bench 2.1 at 67.79 (22.1 under Xiaomi's 89.9), and Terminal-Bench 4.0 at 31.31. The other three agent and coding numbers are Xiaomi's own.
MiMo-V2.6-Pro benchmarks and pricing, every number sourced: 7 tracked MiMo-V2.6-Pro benchmark scores (4 independently run, 3 still resting on a vendor’s own claim), priced at $0.43 per million input tokens and $0.87 per million output.
MiMo-V2.6-Pro architecture: Mixture-of-Experts; 1.02T total parameters (42B activated per token); 1M-token context window.
- Released
- 2026-09-22
- License
- open-weights
- Context window
- 1M tokens
- Knowledge cutoff
- Not disclosed
- Verified
- sources checked 2026-09-22–2026-10-15
- Parameters
- 1.02T (42B active)
- Architecture
- Mixture-of-Experts
MiMo-V2.6-Pro’s verified record
MiMo-V2.6-Pro’s most-compared rival is GLM-5.3-Flash: 2 leads and 5 not callable across their 7 shared comparisons. MiMo-V2.6-Pro is priced at $0.43/$0.87 per 1M tokens (in/out) vs GLM-5.3-Flash’s $0.15/$0.50.
Against the 190 head-to-head comparisons MiMo-V2.6-Pro shares with other tracked models:
A gap counts for MiMo-V2.6-Pro only where both sides were run independently and the benchmark still separates models. Losses are listed alongside wins on purpose — MiMo-V2.6-Pro trails on 24 of them.
HLE · no tools Reasoning
±2 is noiseAnalystAgent Spreadsheet & document analysis
±11.2 is noiseTerminal-Bench 2.1 Terminal ops
±10.6 is noiseTerminal-Bench 4.0 Terminal ops
±12.4 is noiseNo verdict for MiMo-V2.6-Pro anywhere on Agents' Last Exam, DeepSWE, Toolathlon-Verified ( nothing independently confirmed on both sides).
- None of MiMo-V2.6-Pro’s coding comparisons are independently confirmed on both sides yet.
MiMo-V2.6-Pro API pricing
$0.43 in / $0.87 out per 1M tokens — official pricing source
What MiMo-V2.6-Pro costs per job
| Workload | Tokens in / out | Cost |
|---|---|---|
| One long chat turn | 100K / 10K | $0.052 |
| A codebase review | 1,000K / 100K | $0.522 |
| A day of agent work | 10,000K / 1,000K | $5.22 |
Computed from MiMo-V2.6-Pro’s list rates above — cache discounts and batch tiers are not applied.
MiMo-V2.6-Pro is one of 2 Xiaomi models tracked on this site, at these official list prices.
| Model | In / 1M | Out / 1M |
|---|---|---|
| MiMo-V2.6-Flash | $0.14 | $0.28 |
| MiMo-V2.6-Pro | $0.43 | $0.87 |
MiMo-V2.6-Pro benchmark scores
| Benchmark | Score |
|---|---|
| Terminal-Bench 2.1[1] Terminal ops · ±10.6 is noise | |
| DeepSWE[2] Long-horizon coding · ±9.5 is noise | |
| Toolathlon-Verified[3] Multi-tool chores · ±9.7 is noise | |
| Agents' Last Exam[4] Professional work · ±3.2 is noise | |
| HLE(no tools)[5] Reasoning · ±2 is noise | |
| Terminal-Bench 4.0[6] Terminal ops · ±12.4 is noise | |
| AnalystAgent[7] Spreadsheet & document analysis · ±11.2 is noise |
Who ran these numbers: 4 of 7 independent — vals.ai (2), artificialanalysis.ai (2); vendor self-reported (3).
- Terminal-Bench 2.1: vals.ai's own Terminus 2 run, board row 'MiMo V2.6 Pro' (67.79 ± 1.87, no effort parameter listed, pass@1), from its archived Terminal-Bench 2.1 table. Replaces Xiaomi's self-reported 89.9 (model-card table, no effort tier stated), 22.1 points higher. Xiaomi's card reports 89.9 — not stored as a row; this cell holds vals.ai's 67.79.
- DeepSWE: Xiaomi's model-card evaluation table reports 71.9. Xiaomi's announcement instead reports 72.6 as this RL run's after-training endpoint on the same held-out benchmark (up from 58.4 before it) — two official numbers, conflict flagged, neither independently confirmed; deepswe.datacurve.ai did not list this model when checked 2026-09-22.
- Toolathlon-Verified: Xiaomi's model-card evaluation table reports 76.9 for the Pro. toolathlon.xyz's board listed only MiMo V2.5 when checked 2026-09-22.
- Agents' Last Exam: Xiaomi's model-card evaluation table reports 31.6 for the Pro, split not stated. snorkel.ai's board listed only MiMo V2.5 when checked 2026-09-22.
- HLE: Artificial Analysis's own run (text-only 2,158-question subset; the chart data reads 0.4935). AA lists the model as "MiMo-V2.6-Pro" with no effort level, so which reasoning tier produced the score is not stated. The model page on AA shows only the Intelligence Index, not this sub-score.
- Terminal-Bench 4.0: vals.ai board row — mini-swe-agent harness (single bash tool), pass@1 averaged over 3 full passes, raw value 31.313 at no effort tier stated. Artificial Analysis same-conditions run (no tier suffix): 34.85. Xiaomi's card reports 34.9 with no tier stated — not stored as a row; this cell holds vals.ai's independent run.
- AnalystAgent: AA's own run, board row 'MiMo-V2.6-Pro': analystAgent 0.375 = 37.5 pass^5. pass@1 is 61.25. Read from the MiMo-V2.6-Pro page on 2026-10-15.
Notes on the record
Pricing checked 2026-09-22 against Xiaomi's official model page (mimo.mi.com/models/en-US/mimo-v2.6-pro): $0.435 per million input tokens, $0.87 output, $0.0036 cached input — one flat rate with no length threshold, time-of-day discount or promotional condition in the docs. Xiaomi states the V2.6 prices are unchanged from the V2.5 series (announcement page, 2026-09-22).
Spec from the HuggingFace model card (XiaomiMiMo/MiMo-V2.6-Pro-RL, MIT license, uploaded 2026-09-21): sparse MoE, 1.02T total / 42B activated parameters, 1M-token context, 128K max output, text/image/video/audio input. Knowledge cutoff is not disclosed. The official announcement's Update Time reads September 22 while Artificial Analysis and the HuggingFace upload date the release September 21; this site records the official page's date.
Provenance: two tracked scores are independent. One is Artificial Analysis's Humanity's Last Exam run, 49.35 (text-only, no effort tier shown), listed on AA's HLE chart though not on its model page, which shows only the Intelligence Index of 46 (corrected 2026-09-23; this page first said no tracked benchmark had an independent score; pre-v5 reading — Intelligence Index v5 shipped 2026-10-28, and this absolute score is not comparable; this site does not publish MiMo-V2.6-Pro's composite). The other is vals.ai's Terminal-Bench 2.1 run, 67.79 (Terminus 2, no effort parameter listed), which replaced Xiaomi's 89.9 on 2026-10-01. The remaining three coding and agent scores are Xiaomi's own model-card figures; tbench.ai, deepswe.datacurve.ai, toolathlon.xyz, snorkel.ai, livebench.ai, arcprize.org and matharena.ai did not list the model on 2026-09-22. Xiaomi's two official sources disagree on DeepSWE v1.1: the model card's evaluation table says 71.9, while the announcement's training write-up reports 72.6 as the after-training endpoint of this RL run (up from 58.4 before it), on the same held-out benchmark used to argue the training generalizes — a different kind of claim from the card's headline eval, but still Xiaomi's own number either way. The card figure is recorded, the conflict flagged, neither independently confirmed.
The card's 'OSWorld-Verified' 82.0 states no release version — not recordable under version-sliced OSWorld cells.
Compare with
FAQ
Has MiMo-V2.6-Pro been independently benchmarked?
Partially. Artificial Analysis runs it independently: besides a composite Intelligence Index of 46 (pre-v5 reading, checked 2026-09-23; Intelligence Index v5 absolute scores are not comparable — this site does not publish MiMo-V2.6-Pro's composite), its Humanity's Last Exam chart lists the model at 49.35 (text-only, no effort tier shown; checked 2026-09-23). vals.ai's archived Terminal-Bench 2.1 board adds a second independent score, 67.79, and its Terminal-Bench 4.0 run (31.31, default effort) added a third on 2026-10-08. AA-AnalystAgent adds a fourth, 37.5 pass^5, with no effort tier on the Artificial Analysis row (checked 2026-10-15). tbench.ai, deepswe.datacurve.ai, toolathlon.xyz, snorkel.ai, livebench.ai, arcprize.org and matharena.ai did not list the model when checked on 2026-09-22, so the other three coding and agent scores are Xiaomi's own.
How much does the MiMo-V2.6-Pro API cost?
$0.435 per million input tokens and $0.87 per million output tokens on Xiaomi's official pricing page (checked 2026-09-22), with cached input at $0.0036. It is one flat rate — no length threshold, time-of-day discount or promotion — and Xiaomi says it matches the V2.5-series price.
Is MiMo-V2.6-Pro open source?
Yes. The weights are downloadable from HuggingFace (XiaomiMiMo/MiMo-V2.6-Pro-RL) under an MIT license, released together with the technical report, RL environments and training framework per Xiaomi's announcement on 2026-09-22.
How good is MiMo-V2.6-Pro at coding and agent tasks?
Xiaomi's own table reports Terminal-Bench 2.1 at 89.9, DeepSWE v1.1 at 71.9, Toolathlon-Verified at 76.9 and Agents' Last Exam at 31.6. The one that has been re-run did not hold up: vals.ai's independent Terminal-Bench 2.1 run measured 67.79, 22.1 points under the card. The other three have no independent confirmation yet, and Xiaomi's announcement separately reports 72.6 for DeepSWE while describing this RL run's training curve.
Why do two different DeepSWE scores circulate for MiMo-V2.6-Pro?
Xiaomi's HuggingFace model card lists 71.9 in its evaluation table. Its announcement instead reports 72.6 while describing an RL training run: DeepSWE v1.1 improved from 58.4 to 72.6 over the run, offered as evidence the training generalizes beyond its own task mix. Both are Xiaomi's own numbers for what should be the same held-out benchmark. This site records the model-card figure (71.9) and flags the conflict; neither has been independently confirmed (checked 2026-09-22).
What is MiMo-V2.6-Pro's best verified score?
Terminal-Bench 2.1 at 67.79 (2026-10-01), the highest independently-confirmed figure on a current board. The DeepSWE rows in its record are the two different official-board scores — which of them to read is untangled in a separate question above.
Further reading
- Models with 10M token context windows 2026 — MiMo-V2.6-Pro is one of the 47 models it compares.
Benchmark guides
- Terminal-Bench 2.1 — what MiMo-V2.6-Pro’s agentic score on it does and doesn’t prove.
- DeepSWE — what MiMo-V2.6-Pro’s coding score on it does and doesn’t prove.