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

Flash-RL scores 38 on the Artificial Analysis Intelligence Index, well above the open-weight median (18) though behind Pro-RL's 46.32. On agentic evals Xiaomi reports directly, it scores 87.6 on Terminal-Bench 2.1 and 80.8 on OSWorld-Verified, both close to Pro-RL's 89.9 and 82.0 and to the closed-source Claude Opus 5 (89.1, 83.4) named in the shared technical report. It trails badly on cybersecurity exploit evals: 25.3 on ExploitBench against Claude Opus 5's 70.0.

The model ships the same 1 million token context window as Pro-RL, with a 128,000 token max output and an identical 5-layer speculative decoder design. Artificial Analysis measured 59.2 output tokens per second, faster than Pro-RL's 48.6 but still under the open-weight class median of 85.7 t/s, plus a 3.80-second time-to-first-answer-token. Flash-RL was also the more verbose of the two on the Intelligence Index, generating 240 million output tokens against Pro-RL's 140 million.

Xiaomi's own product page and this model's Hugging Face card both list Flash-RL's input modality as text, image, video, and audio, matching Pro-RL. Artificial Analysis, which tests models directly rather than citing vendor claims, found only text and image input actually working through Flash-RL's API as of this listing. Output remains text only in every source.

Xiaomi's API prices Flash-RL far below Pro-RL: $0.14 per 1M input tokens on a cache miss, $0.0028 per 1M on a cache hit (about a 98% discount), and $0.28 per 1M output tokens. Artificial Analysis computes a blended rate of $0.06 per 1M tokens, a third of Pro-RL's $0.18.

Weights are published on Hugging Face and ModelScope under MIT. Self-hosting recipes are lighter than Pro-RL's: SGLang with tensor-parallel 8 and data-parallel 2, or vLLM with tensor-parallel-size 4, roughly half the GPU count Pro-RL calls for. Distribution is narrower than Pro-RL: only 1 API provider was live on Artificial Analysis's tracker at listing time, versus 3 for Pro-RL; access otherwise runs through the same first-party channels (Xiaomi's own API, AI Studio, MiMo Desktop, OpenRouter).

As with Pro-RL, Xiaomi's Trust Center had not published a MiMo-V2.6-specific training-data summary or system card at listing time. The shared technical report describes the same "Aligned RL" self-correction stage and adversarial/verifier hardening used to train both checkpoints together.

Flash-RL suits high-frequency, cost-sensitive workloads needing agentic and coding competence close to Pro-RL's, and teams that can accept its narrower API distribution or self-host on roughly half the GPU budget. It is weaker for tasks needing Pro-RL's top intelligence score, offensive-security work, or workflows relying on the vendor's claimed video/audio input, which Artificial Analysis could not confirm.

Flash-RL trained alongside Pro-RL in the same reinforcement-learning campaign: Xiaomi reports its run also completed 30 GRPO steps in under six days, sharing the roughly 750,000-trajectory batch structure, at a reported cost of $850,000, well below Pro-RL's $2.62 million.

There is no MiMo-V2.5-Flash predecessor in Xiaomi's comparison table, so Flash-RL's closest lineage comparison is MiMo-V2.5-Pro, which it beats on every shared benchmark, for example 87.6 vs 65.2 on Terminal-Bench 2.1. Xiaomi has already previewed MiMo-V3, reported to adopt a new "HySparse2" architecture to cut long-context serving costs, as the next model in the line. A smaller research sibling, MiMo-V2.6-Distill-Qwen-9B, was released alongside it for self-hosting on a single GPU.