MiMo-V2.6-Flash: 15B Active Parameters, Close Behind Pro

The interesting question about a small model in a two-model family is not how good it is in isolation. It is where, specifically, it falls behind the large one — because that tells you which tasks to route up and which to keep cheap. MiMo-V2.6-Flash is unusually easy to answer that question for, because Xiaomi published both models in the same benchmark table.

The short version: on most of Xiaomi's agentic coding and tool-use benchmarks, Flash sits within a few points of V2.6-Pro. On a handful of tasks the gap is wide, and those tasks have something in common.

What Flash is

MiMo-V2.6-Flash was released in September 2026 alongside V2.6-Pro, with weights on Hugging Face under the MIT licence. The model card describes:

• A sparse mixture-of-experts model with 309B total and 15B active parameters (Hugging Face's metadata rounds it to 311B).
• 48 layers, 39 sliding-window and 9 global attention, with 256 routed experts and 8 active per token.
• Native text, image, video and audio input via a 681M-parameter vision encoder and a separate audio tokenizer and encoder.
• A 1M-token context window.
• A five-layer multi-token prediction drafter that proposes seven tokens per forward pass for parallel verification — the mechanism behind speculative decoding, covered in speculative decoding explained.

The architecture is in the same class as V2.5. The change is training. Xiaomi's technical report, "MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement", describes scaling RL compute across larger batches, more diverse code, general, visual and cyber environments, and more grader compute, with grading designed to push the model towards shorter, more token-efficient solutions.

The Flash-to-Pro gap, benchmark by benchmark

From the V2.6 model card, Flash against Pro:

• Terminal-Bench 2.1 — 87.6 against 89.9.
• DeepSWE v1.1 — 67.9 against 71.9.
• OSWorld-Verified — 80.8 against 82.0.
• Toolathlon-Verified — 73.6 against 76.9.
• AutomationBench — 52.3 against 53.1.
• ProgramBench — 26.0 against 26.5.
• CyberGym — 95.1 against 94.0, the one place Flash leads.

For everyday agentic coding, terminal work and computer use, those are small gaps. OpenRouter listed Flash at $0.14 input and $0.28 output per million tokens at the time of writing, against $0.435 and $0.87 for Pro. Roughly a third of the price for most of the capability is a strong position.

Where the gap is real

The wide gaps cluster in two places:

• The hardest terminal tasks. Terminal-Bench 4.0 scores 28.8 for Flash against 34.9 for Pro.
• Exploitation and security research. ExploitGym 6.0 against 17.8, ExploitBench 25.3 against 47.9, SEC Bench Pro 47.5 against 66.3.

The common thread is long chains of precise reasoning where one wrong step ruins the result. Finding a vulnerability and building a working exploit is about as unforgiving as a task gets. Flash handles the breadth; Pro handles the depth. Xiaomi also reports no GDPval-AA score for Flash, so if professional document work is your use case, you have no vendor signal and need your own test.

If your workload looks like that, the MiMo-V2.6-Pro guide is the better starting point.

RL or MOPD: picking the open checkpoint

If you self-host, Hugging Face has two Flash checkpoints. MiMo-V2.6-Flash-RL is the output of the RL stage. MiMo-V2.6-Flash-MOPD is described as the MOPD upgrade of that checkpoint: a further pass that distils several domain-specialised teachers into the model on-policy, including a short run aimed at reducing tool-call repetition.

Xiaomi does not say outright which one to deploy. For agent loops, repeated tool calls are one of the more expensive failure modes, so the MOPD checkpoint is the sensible default unless your own tests show otherwise. Xiaomi's vLLM example uses four-way tensor parallelism and warns that stable vLLM releases may lag, pointing to a dedicated image instead.

There is also MiMo-V2.6-Distill-Qwen-9B on the same organisation page, a much smaller distilled model, for anyone who needs something that runs on a single modest GPU. It is a different model and should be evaluated separately.

Flash against V2.5

The MiMo-V2.5 guide covers the previous model in this size class. Both are omnimodal, both have 1M context, and both were listed at the same OpenRouter price. Xiaomi did not publish a direct Flash-versus-V2.5 comparison, but Flash beats V2.5-Pro by wide margins on the V2.6 table — 87.6 against 65.2 on Terminal-Bench 2.1, for instance — so it would be surprising if it did not beat the smaller V2.5 as well.

Route Flash first

The benchmark profile suggests a clean routing setup: send everything to Flash, escalate to Pro when the task is security-sensitive, when a run fails its checks, or when the job is known to need very long precise chains. Model routing and fallbacks covers how to implement the escalation without doubling latency on the common path.

To confirm the split for your own workload, take twenty real tasks, run both models on the same harness, and compare cost per completed task. If Flash completes nearly as many at a third of the price, the routing decision is made for you.