Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.
Official5
mimo.xiaomi.com2026-09-21officialHN front pageabout the MiMo V2 familyHN 1,130 pts · 483 comments3,621 words
Xiaomi releases the MiMo-V2.6 series on September 22nd, 2026, introducing MiMo-V2.6-Pro, MiMo-V2.6-Flash, and the new MiMo-V2.6-Pro-UltraSpeed variant delivering up to 20x faster output speed at the same quality. The models completed 30 RL steps over roughly 750k trajectories at costs of about $0.85M and $2.62M respectively, with DeepSWE v1.1 scores rising from 48.8 to 65.68 for the Pro model. This launch marks a key step in scaling RL compute on verifiable, complex tasks.written, reviewed, cited
- MiMo-V2.6-Pro-UltraSpeed delivers up to 20x faster output speed at the same quality for users requiring extreme generation speed.
- The training run completed in under six days with MiMo-V2.6-Pro costing about $2.62M and MiMo-V2.6-Flash costing about $0.85M.
mimo.xiaomi.com2026-07-07officialHN front pageabout the MiMo V2 familyHN 116 pts · 49 comments4,318 words
Xiaomi engineers optimized the MiMo-V2.5 series for inference by implementing a tiered caching system that supports Hybrid Sliding Window Attention. This architecture reduces KVCache storage to roughly 1/7 of Full Attention, improving overall KVCache capacity efficiency by approximately 7×. The system uses SGLang v0.5.5 as the serving backend and resolves storage conflicts by maintaining separate Full KV and SWA KV pools to strictly constrain SWA storage to O(W).written, reviewed, cited
- Medium and short sequences retain relatively dense SWA KV Cache at fixed length positions based on user request patterns.
mimo.xiaomi.com2026-09-27officialabout the MiMo V2 familyHN 2 pts · 1 comments1,793 words
MiMo-V2.6 models exhibited tool-call repetition where agents issued identical calls repeatedly, wasting context and stalling tasks. The team defined exact within-turn repetition as (N-U)/N and distinguished it from flooding, then adopted an MOPD-based approach to fix the issue. This solution cost approximately $90,000, significantly less than the $2.31 million estimated for the penalty-threshold fix, and reduced repetition rates substantially.written, reviewed, cited
- The flooding penalty threshold was lowered from 32 calls to eight to suppress abnormal call volume.
- Applying the stricter flooding penalty at full scale required restarting 20 training steps at an estimated cost of $2.31 million.
mimo.xiaomi.com2026-04-22officialabout the MiMo V2 familyHN 24 pts · 3 comments1,045 words
Xiaomi releases MiMo-V2.5-Pro, a 1.02T-parameter model with 42B active parameters and a 1M-token context window built on hybrid-attention architecture. The model autonomously implemented a SysY compiler in 4.3 hours across 672 tool calls, scoring 233/233, and generated an 8,192-line video editor in 11.5 hours using 1,868 tool calls. It achieves 64% Pass^3 on ClawEval using only ~70K tokens per trajectory, representing roughly 40–60% fewer tokens than comparable frontier models.written, reviewed, cited
- The model built the compiler layer by layer, perfecting Koopa IR (110/110) before the RISC-V backend (103/103).
- A trained analog designer typically spends several days on a project of this scope involving six simultaneous metrics.
- Pre-training runs on 27T tokens using FP8 mixed precision at a native 32K sequence length.
mimo.xiaomi.com2026-04-25officialabout the MiMo V2 familyHN 3 pts · 0 comments520 words
Analysis5
artificialanalysis.ai2026-09-22authorityHN front pageabout the MiMo V2 familyHN 166 pts · 68 comments973 words
MiMo-V2.6-Pro is a multimodal open-weight model from Xiaomi released on September 21, 2026, that scores 46 on the Artificial Analysis Intelligence Index. It costs $0.43 per 1M input tokens and $0.87 per 1M output tokens while generating output at 52 tokens per second. The model operates under the MIT license with 1.0 trillion total parameters and 42 billion active parameters during inference.written, reviewed, cited
- The model uses a 7:2:1 cache hit/input/output ratio to calculate a blended rate of $0.18 per 1M tokens.
- The model generates 140M output tokens on the Intelligence Index, which is described as somewhat higher than average.
sebastianraschka.com2026-09-22authorityabout the MiMo V2 familyHN 3 pts · 1 comments202 words
The model uses Grouped Query Attention with Sliding Window Attention at a 128-token window size. DeepSWE pass@1 accuracy on held-out harnesses improved from approximately 50% to 66%. Training employed large RL batches of 1,568 prompts with 16 rollouts generating 25,088 trajectories.generated, cited
artificialanalysis.ai2026-09-27authorityabout the MiMo V2 familyHN 2 pts · 1 comments961 words
MiMo-V2.6-Flash is a 309 billion parameter Mixture of Experts reasoning model released by Xiaomi on September 21, 2026, under the MIT license. It scores 38 on the Artificial Analysis Intelligence Index but operates at 60 tokens per second with a 4.42s time to first token. The model supports text and image input with a 1M token context window and is available through 4 API providers.written, reviewed, cited
- The model uses 15 billion active parameters during inference despite having 309 billion total parameters.
- MiMo-V2.6-Flash generated 240M output tokens during evaluation, which is significantly higher than the median of 140M.
firethering.com2026-04-28about the MiMo V2 familyHN 21 pts · 7 comments1,797 words
Xiaomi released MiMo-V2.5-Pro, an open-source coding model that finished a university compiler project in 4.3 hours and built a video editor using 1,868 tool calls. The model utilizes a hybrid attention design to maintain functionality at 1M tokens and activates 42B parameters from a 1.02T total parameter count. Distributed under an MIT license, it offers token efficiency advantages over closed models for agentic workflows.written, reviewed, cited
- MiMo-V2.5-Pro recovers from a regression at turn 512 without external intervention during long-horizon tasks.
- The model scores 0.37 and 0.62 on GraphWalks subtasks at the full 1M token length.
forkast.news2026-09-18about the MiMo V2 familyHN 3 pts · 0 comments643 words
