This morning’s Lessengers 3.2T glass NPO note was a package next to the switch ASIC CapEx sentence. Tonight is the other half of the stack again: what you route agent traffic to once the hall is paid for — and whether that route has to stay closed. On September 21, 2026 (September 22 in Beijing / Japan coverage), Xiaomi’s MiMo team shipped MiMo-V2.6-Pro and MiMo-V2.6-Flash under an MIT license with weights on Hugging Face, the same day SpaceXAI shipped closed Grok 4.7.
Primary sources: MiMo-V2.6-Pro-RL model card, MiMo-V2.6-Flash-RL, Xiaomi MiMo V2.6 page, Artificial Analysis model page, DataNorth Sep 22 read, Kingy.ai Sep 21–22 specs + AA table, Gigazine Sep 24, and Grok 4.7 launch / AA’s Grok 4.7 note for the same-day foil. Vendor benches below are Xiaomi’s unless marked independent (Artificial Analysis).
Spec sheet (named numbers)
| MiMo-V2.6-Pro | MiMo-V2.6-Flash | |
|---|---|---|
| HF checkpoint | XiaomiMiMo/MiMo-V2.6-Pro-RL |
XiaomiMiMo/MiMo-V2.6-Flash-RL |
| API id (common) | mimo-v2.6-pro (OpenRouter xiaomi/mimo-v2.6-pro) |
mimo-v2.6-flash |
| Architecture | Sparse MoE | Sparse MoE |
| Total / active params | 1.02T / 42B | 309B / 15B |
| Context / max output | 1,048,576 / 131,072 (OpenRouter / Kingy) | Same context |
| Backbone (Pro) | 70 layers (60 SWA / 10 GA), hidden 6144, 384 routed experts / 8 active, SWA window 128 | 48 layers (39 SWA / 9 GA), hidden 4096, 256 / 8 |
| Vision | Shared 681M MiMo ViT (28 layers) | Same |
| Audio | 308M AudioTokenizer + 127M patch encoder | Same |
| Speculative decode | 5-layer SWA MTP (predicts 7 tokens / forward) | Same |
| Modalities | Text, image, video, audio in; text out | Same |
| License | MIT (commercial self-host allowed) | Same |
| List price (per 1MTok) | $0.435 in / $0.87 out | $0.14 / $0.28 |
| Cached input | $0.0036 | $0.0028 (Gigazine) |
| UltraSpeed (Pro only) | $4.35 / $8.70 (10× list; vendor “up to 20×” speed) | — |
| Batch API | 50% of Pro standard (Kingy) | — |
| AA Intelligence Index | 46 (#1 open weights) | — |
| AA cost / index task | ~$0.13 | — |
| Hosted | Xiaomi API / AI Studio / MiMo Code / Desktop + OpenRouter | Same |
| Self-host cookbook | SGLang multi-node (tp 16, 2 nodes in card); vLLM recipe |
SGLang tp 8; vLLM tp 4 |
Same-day foil: Grok 4.7 also scores 46 on AA at $2 / $6, 500k context, closed weights (
xAI launch
). That is the price-floor sentence Xiaomi is selling — not a claim that Pro beats
Claude Opus 5.5
or
GPT-6 Astra
on every long-horizon board.
What actually changed for people who run agents
The durable product change is not a new FLOPS sticker. It is MIT weights + a rate card that sits under Grok’s same-day AA score.
Artificial Analysis puts Pro at 46 on Intelligence Index v4.3.x — top open-weights row in Kingy’s Sep 21 table — at about $0.13 per index task versus Grok 4.7’s ~$3.74 at xhigh and Claude Opus 5 max at ~$5.86 on that same secondary table. List tokens are $0.435 / $0.87 with a deep cache-read discount ($0.0036). Flash cuts that to $0.14 / $0.28 at 15B active. UltraSpeed is a latency SKU at 10× Pro list, not a smarter checkpoint.
Architecture explains the serving math: only 42B of 1.02T fire per token. That does not mean you can park Pro on a spare workstation — the published SGLang command is a two-node bring-up. Flash’s tp 8 card is the realistic self-host on-ramp.
Vendor benches (Xiaomi’s table) vs the independent index
Xiaomi’s Hugging Face evaluation table (selected rows; higher better):
| Benchmark | MiMo Pro | MiMo Flash | Claude Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| DeepSWE v1.1 | 71.9 | 67.9 | 74.0 | 73.0 |
| AutomationBench v1.0.6 | 53.1 | 52.3 | 50.3 | 45.8 |
| Terminal-Bench 2.1 | 89.9 | 87.6 | 89.1 | 88.8 |
| Terminal-Bench 4.0 | 34.9 | 28.8 | 49.0 | 39.9 |
| OSWorld-Verified | 82.0 | 80.8 | 83.4 | 83.0 |
| Agents’ Last Exam | 31.6 | 27.6 | 31.6 | 30.8 |
| GDPval-AA 2.1 (Elo) | 1673 | — | 1708 | 1588 |
| ExploitBench | 47.9 | 25.3 | 70.0 | 78.5 |
| CyberGym | 94.0 | 95.1 | — | — |
Pattern that belongs in a routing ticket: Pro is competitive on general agent / tool-use rows Xiaomi highlights, and it falls off on harder long-horizon coding and exploit boards — especially Terminal-Bench 4.0 (34.9 vs Opus 5 49.0; Kingy also cites Astra 59.6 on Xiaomi’s wider appendix). CyberGym 94 with MIT weights is a dual-use note, not a marketing flourish. Treat every vendor row as Xiaomi-harness until your suite says otherwise; some secondaries also flag small number drift between Xiaomi pages (DeepSWE 71.9 vs RL-section 72.57).
Independent check: AA Index 46 is the outside number. AA’s model page also flags verbosity (~140M output tokens across the index run) and lists ~44.4 output tok/s in its open-weights class comparison — Kingy earlier cited ~130 tok/s on Xiaomi’s API. Measure the endpoint you will actually bill.
Where it sits next to Sol/Luna, Opus 5.5, and this morning’s optics
| Purchase | Product | What it buys |
|---|---|---|
| Flagship reasoning (OpenAI) | GPT-6 Astra | Highest OpenAI ceilings; $10/$50 |
| Default coding agents (OpenAI cost ladder) | GPT-6 Sol | $2/$10; cache-steerable effort |
| Volume / extract (OpenAI) | GPT-6 Luna | $0.10/$0.50 |
| Flagship / agentic coding (Anthropic) | Claude Opus 5.5 | Fable-class work at $4/$20; fallback safeguards |
| Open-weights price floor | MiMo-V2.6-Pro | AA 46; $0.435/$0.87; MIT; 1.02T/42B |
| Open volume / self-host on-ramp | MiMo-V2.6-Flash | $0.14/$0.28; 309B/15B; tp 8 cookbook |
| Same-day closed peer | Grok 4.7 | AA 46; $2/$6; closed; 500k context |
| Glass NPO package (this morning) | Lessengers 3.2T DOW | Millimeters-from-ASIC optics |
MiMo does not move KV blocks and does not invent a fabric. It argues that a large slice of agent loops should not burn Astra or Opus tokens, and that the open-weights invoice finally has a named MoE sheet at the same AA Index point as Grok 4.7.
Caveats I would write into any MiMo ticket
- Reproduce Terminal-Bench 4.0 / DeepSWE / your repo harness before you price a SKU — Xiaomi’s table is not your fleet.
- Do not paste “on par with Opus 5.” Xiaomi’s own rows trail on most shared Opus 5 comparisons; the honest win is cost at AA 46, not a blanket ceiling claim. Opus 5.5 is also newer than the Opus 5 column in Xiaomi’s card.
- Self-host Pro ≠ free. Weights are MIT; serving still needs a multi-GPU / multi-node bill. Flash first unless you have the cluster.
- UltraSpeed is 10× price. Latency tool — measure tok/s on the endpoint you bill.
- Cyber capability is open. CyberGym 94 under MIT means strong vuln-analysis weights without an API gate — design your own controls.
- Not a rematch of Lessengers / PicoJool / Continuum. Those are optics and campus fabric. This is the model routing layer.
How I would evaluate if I ran agent fleets
-
Cost per completed task on Flash and Pro vs
Sol,
Luna,
and
Opus 5.5
— not just
$/MTok. - Terminal-Bench 4.0-class unattended jobs where Xiaomi’s own gap is widest; keep Astra/Opus on a shadow route.
-
Cache-hit and verbosity on your prompts — AA’s cheap
$/task assumes its token mix. - Self-host only after Flash proves quality on your suite; escalate to Pro weights when the cluster math clears.
- Guardrails in code for unattended cyber-capable open weights, not prompt warnings alone.
This is an open-weights price-floor release at Artificial Analysis 46, not a new closed flagship. Route hard unattended jobs to Astra or Opus 5.5 until your harness says otherwise. Put Flash on volume and Pro on the harder open tier when $0.13-class index economics beat the Terminal-Bench 4.0 risk. Measure, then move traffic.
Create-only. Specs and dates from Xiaomi’s MiMo-V2.6 Hugging Face cards, the MiMo V2.6 site, Artificial Analysis’s MiMo-V2.6-Pro page, DataNorth / Kingy.ai / Gigazine secondary reads, and xAI / AA for same-day Grok 4.7. Update when a named harness reproduces Terminal-Bench 4.0 / DeepSWE against Opus 5.5 and Astra, or when Xiaomi publishes a fuller system card.
Sources: MiMo-V2.6-Pro-RL · MiMo-V2.6-Flash-RL · Xiaomi MiMo V2.6 · Artificial Analysis · DataNorth · Kingy.ai · Gigazine · xAI Grok 4.7 · AA Grok 4.7 note.
Anpoo Sivanadi, Staff Software Engineer