MiMo-V2.6-Pro: What Scaled RL Changed in Xiaomi's Flagship

Many model upgrades change the architecture or the parameter count and let the benchmarks follow. MiMo-V2.6-Pro kept the size of its predecessor — 1.02T total parameters, 42B active — and changed the training. The result, on Xiaomi's own numbers, is a very large generation-on-generation jump on agentic benchmarks.

That makes it a useful case study in how far post-training alone can move a model, and also a reason to read the numbers with care. Here is what changed, what the benchmark table does and does not show, and how to check the claims on your own work.

Same size, new inputs, new training

V2.6-Pro was released in September 2026 with weights on Hugging Face under the MIT licence. The backbone matches V2.5-Pro: 70 layers, 60 sliding-window and 10 global attention, 384 routed experts with 8 active per token, and a 1M-token context window.

Two things are different. First, it is now omnimodal. V2.5-Pro was text-only; V2.6-Pro adds a 681M-parameter vision encoder plus an audio tokenizer and encoder, so images, video and audio are native inputs. Second, the multi-token prediction drafter grew to five layers that propose seven tokens per forward pass, which matters for serving speed.

The training story is in Xiaomi's technical report, "MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement". The team scaled RL along three axes: bigger asynchronous batches with contexts up to 1M, more varied environments across code, general, visual and cyber tasks, and more compute spent on grading long-horizon runs. They froze the MoE router for stability and added defences against reward hacking. A final MOPD2 stage distils several specialised teachers into the model on-policy.

The numbers against V2.5-Pro

The V2.6 model card puts both generations side by side:

• Terminal-Bench 2.1 — 89.9, up from 65.2.
• DeepSWE v1.1 — 71.9, up from 19.0.
• Toolathlon-Verified — 76.9, up from 49.1.
• AutomationBench — 53.1, up from 16.0.
• GDPval-AA 2.1 — 1673, up from 1107.
• Agents' Last Exam — 31.6, up from 13.2.
• OSWorld-Verified — 82.0, with no V2.5-Pro score since that model could not see a screen.

Every row in that comparison is a long, multi-step agentic task, which is exactly what the RL environments targeted. The table does not report knowledge or single-turn reasoning benchmarks for either generation, so it tells you nothing about whether those moved.

Reading a vendor table carefully

Three cautions before you quote any of this.

The harness is Xiaomi's. Agentic scores depend heavily on the scaffold, the tool definitions and the turn budget. A vendor running its own model on its own harness is the best case. Terminal-Bench explained covers why the same model can move several points between harnesses.

Versions matter. The card reports GDPval-AA 2.1. Other published leaderboards use GDPval-AA v2. Do not line up scores from different versions as if they were the same scale.

Some benchmarks are in-house. MiMo Code Bench, MiMo Cyber Bench and MiMo VisualCoding are Xiaomi's own. They may be well designed, but nobody outside can check them yet. Weight the public ones more heavily.

None of that makes a 24-point gain on Terminal-Bench 2.1 disappear. It just means the honest summary is "a large improvement, size to be confirmed on your work".

The security results deserve attention

V2.6-Pro scores 94.0 on CyberGym, 47.9 on ExploitBench and 66.3 on SEC Bench Pro, against 40.0, 16.6 and 17.7 for V2.5-Pro. Those benchmarks measure vulnerability discovery and exploitation. That is valuable for defensive teams running audits, and it is also a capability you should assume any agent built on this model has.

If you give it shell access, network access or credentials, treat it as a capable operator rather than a code completer. Agent sandboxing covers the isolation that should already be in place for any agent; with this model it moves from good practice to requirement.

Pro, Flash or UltraSpeed

OpenRouter listed V2.6-Pro at $0.435 input and $0.87 output per million tokens at the time of writing — the same as V2.5-Pro, so the upgrade costs nothing per token. There are two alternatives in the same generation:

• V2.6-Flash is about a third of the price and within a few points on most benchmarks, falling well behind mainly on exploitation tasks and the hardest terminal work. The MiMo-V2.6-Flash guide has the full comparison.
• V2.6-Pro-UltraSpeed is a faster edition of this model at ten times the price. The UltraSpeed guide covers when that is worth paying.

Self-hosters get two checkpoints, MiMo-V2.6-Pro-RL and MiMo-V2.6-Pro-MOPD. The MOPD version is the later stage and includes a fix for tool-call repetition, so it is the natural default for agents. Xiaomi's SGLang example spans two nodes with 16-way tensor and expert parallelism; this is a cluster deployment, not a workstation one.

What to test first

1. If you run V2.5-Pro today, replay twenty completed tasks on V2.6-Pro with the same harness and compare turns to completion and merge rate.
2. If you route between Flash and Pro, measure where Flash fails on your tasks, not on Xiaomi's.
3. If you planned multimodal work with a separate vision model, test whether V2.6-Pro alone can replace the pair.
4. Track output tokens per task. Xiaomi says its grading rewards shorter solutions; check whether that holds in your loop, since it changes the real cost.