MiMo-V2.6-Flash review, pricing and limits

For teams that want to self-host an open-weight agentic model without per-token fees, MiMo-V2.6-Flash keeps inference cheap by activating only a small fraction of its total parameters per token, though every benchmark Xiaomi has published so far is its own, unverified by outside evaluators.

MiMo-V2.6-Flash is Xiaomi's MIT-licensed sparse Mixture-of-Experts model with a 1M-token context window, launched in September 2026 with native text, image, video and audio input understanding, positioned as the efficient, cheaper tier beneath the larger MiMo-V2.6-Pro model announced the same week.

Provider: Xiaomi · Family: MiMo V2.6

More about Xiaomi on HokAI

Context window: 1,048,576 tokens · Max output: 131,072

Input modalities: text, image, video, audio, tool-calls · Output: text, tool-calls

About MiMo-V2.6-Flash

MiMo-V2.6-Flash is Xiaomi's efficient-tier model in the MiMo-V2.6 family, built by the MiMo research team under Luo Fuli, a researcher who joined Xiaomi from DeepSeek in late 2025. Xiaomi announced the model alongside the larger MiMo-V2.6-Pro on September 21-22, 2026, following a public reinforcement-learning training run that began September 15, 2026 and streamed real-time cost, token throughput and per-domain reward metrics on a public dashboard (mimo.xiaomi.com/rl), a departure from the closed system-card practice of most Western labs. Flash uses a sparse Mixture-of-Experts architecture with 309 billion total parameters and 15 billion activated per token across 256 routed experts (8 active per token), a hybrid of 39 sliding-window and 9 global-attention layers, and ships under the MIT license.

On specifications, Flash carries a native 1,048,576-token context window and a 131,072-token maximum output, confirmed independently on both Vercel's AI Gateway listing and llm-stats.com. It takes text, image, video and audio input through a 681-million-parameter vision encoder and a dedicated audio tokenizer, and Xiaomi's own MiMo platform lists tool calling, JSON mode, web search and a deep-thinking reasoning mode among its supported features. A multi-token-prediction speculative decoder is included, which Xiaomi says delivers 2.5 to 3.7 times faster inference than standard autoregressive decoding, though that figure is a vendor claim rather than an independently measured one.

On performance, Xiaomi has not published any standard third-party benchmark for Flash: no MMLU, GPQA, ARC-AGI, HumanEval or SWE-bench Verified score appears on its model card. The only reported figures are Xiaomi's own custom evaluations: 67.9 on its DeepSWE v1.1 coding benchmark, 26.0 on ProgramBench, 61.2 on MiMo Code Bench, 73.6 on Toolathlon-Verified and 80.8 on OSWorld-Verified. Independent write-ups (OrcaRouter, September 2026) note explicitly that none of these scores has been independently reproduced or appears on a public leaderboard, and flag documentation inconsistencies in Xiaomi's own materials, including conflicting parameter-count figures and a mismatch between the stated 5-layer and the configured 3-layer speculative-decoding drafter. Separately, the larger MiMo-V2.6-Pro variant debuted as the top-scoring open-weights model on the Artificial Analysis Intelligence Index (score 46) as of September 2026, per Artificial Analysis's own announcement, though that score applies to Pro, not to Flash.

Pricing is confirmed directly from Xiaomi's MiMo platform and cross-checked on Vercel's AI Gateway, with a steep discount for cached input versus a cache miss and a separate, higher rate for output tokens (see pricing below). The API is documented as compatible with both the OpenAI and Anthropic request protocols, so existing SDK integrations built for either provider need minimal changes to point at Xiaomi's endpoint. Because the weights are MIT-licensed, teams can also download and self-host Flash at no licensing cost, though the FP8 release weighs roughly 172.9 gigabytes across 65 shards and Xiaomi's own deployment guidance calls for a multi-GPU node (SGLang with tensor-parallel 16 and data-parallel 2, or vLLM with tensor-parallel 8) rather than a single GPU.

On safety, Xiaomi describes an aligned reinforcement-learning process with a cold start from self-correction, where the model reflects on and rewrites its own misaligned outputs, plus environment hardening and verifier cross-checks during RL to limit reward hacking. No formal system card, named red-team partners, or responsible-scaling-style framework has been published for the MiMo-V2.6 family. Independent reporting from Forkast (September 2026) also flagged an unexplained "Claude Distill Requests" line item inside Xiaomi's own public training dashboard, raising an open question about whether a third-party proprietary model played an undisclosed role in training, which Xiaomi has not addressed publicly.

Flash fits builders comfortable acting on Xiaomi's own claims ahead of outside confirmation, in exchange for a big context window and a permissive license. It is the wrong pick for procurement that requires SOC 2 or HIPAA-documented handling, single-GPU or edge deployment, or third-party-verified scores; those buyers are better served by an already-audited alternative such as a frontier US lab's model or a DeepSeek or Qwen release with public leaderboard results.

Pricing

Xiaomi's MiMo platform lists a cache-miss rate of $0.14 per 1M input tokens, a cache-hit rate of $0.0028 per 1M tokens, and $0.28 per 1M output tokens; Vercel's AI Gateway shows matching figures (September 2026). Because the weights are MIT-licensed, self-hosting avoids these API rates entirely at the cost of your own GPU infrastructure.

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• 1M-Token Native Context: Handles up to 1,048,576 input tokens and 131,072 output tokens in a single call, confirmed on Vercel AI Gateway and llm-stats.com listings (Sep 2026).
• Omnimodal Input Understanding: Reads text, images, video and audio via a 681M-parameter vision encoder and a dedicated audio tokenizer, per the Hugging Face model card.
• Sparse MoE Efficiency: Routes each token through 8 of 256 available experts instead of the full network, the sparse-MoE design behind the model's lower per-token serving cost.
• Prompt Caching: Cached input tokens are billed at a steep discount versus a cache miss, cutting costs for agent loops that repeat the same context turn after turn (see pricing for exact rates).
• Speculative Decoding: Ships with a multi-token-prediction drafter that Xiaomi says delivers 2.5 to 3.7x faster inference than standard autoregressive decoding (vendor-reported, Sep 2026).

Pros

• MIT-licensed and self-hostable, activating only a small share of its total parameters per token so teams can run it on owned multi-GPU infrastructure without per-token vendor fees.
• A 1M-token context window with a large max output, among the bigger publicly documented context windows in an open-weights model as of September 2026.
• Steeply discounted cache-hit pricing on repeated context makes long-running agent workloads and large codebases unusually cheap next to the standard cache-miss rate.

Cons

• No standard third-party benchmark (MMLU, GPQA, SWE-bench Verified, ARC-AGI) is published; every reported score is Xiaomi's own custom benchmark, unverified on any public leaderboard at release.
• Full FP8 weights run about 172.9GB across 65 shards and need a multi-GPU node (SGLang TP16/DP2 or vLLM TP8) to serve, ruling out single-GPU self-hosting.
• A third-party report (Forkast) spotted a 'Claude Distill Requests' entry buried in Xiaomi's own live training dashboard, an unresolved question about undisclosed reliance on another vendor's model that Xiaomi has not commented on.

Visit MiMo-V2.6-Flash Official Page