MiMo-V2.6 Technical Deep Dive

On September 22, 2026, Xiaomi released and open-sourced the MiMo-V2.6 series. This article is compiled from the official launch page and its full benchmark appendix, and focuses on three things: the complete benchmark data and the limits of its comparability, the training curves and engineering behind scaled reinforcement learning, and a pricing-based model selection analysis. Application demos and research case studies are summarized only briefly.

---

1. Release Highlights

• Two natively omnimodal models: MiMo-V2.6-Pro (flagship) and MiMo-V2.6-Flash (efficiency tier), plus MiMo-V2.6-Pro-UltraSpeed, a fast mode of Pro with up to 20x output speed.
• Highest-scoring open-source model: Pro scores 46.32 on the Artificial Analysis Intelligence Index v4.3 (September 2026), surpassing Kimi K3 and Qwen3.8 Max.
• Unchanged pricing: API pricing is carried over from the V2.5 series.
• RL run streamed live: under six days, 30 RL steps, roughly 750k trajectories, with the production run streamed publicly. The technical report, training environments, and RL code are all open-sourced.
• Official tagline: Frontier intelligence, all the modalities, built in public.

---

2. Model Lineup and Pricing

USD per million tokens. Cache writes are free for a limited time. A Token Plan is available for high-volume usage. RMB pricing is listed on the official pricing page.

A few observations:

• Flash's output price is roughly one-third of Pro's, while the gap on most benchmarks is only 1 to 4 points (see below). It is the obvious value option.
• UltraSpeed costs 10x Pro for up to 20x speed. It suits latency-sensitive interactive use, not batch or offline workloads.
• Cache-hit pricing is very low ($0.0036 for Pro). For agent applications with long system prompts and frequent repeated calls, effective cost will be far below list price.

---

3. Benchmark Overview

Data below comes from the appendix on the official launch page (the full comparison set). "Pro rank" is computed among models with a reported score in the appendix; ties are noted.

3.1 Coding

Coding is where the generation gain is largest. DeepSWE v1.1 went from 19.0 to 71.9, essentially from unusable to first tier, and ProgramBench doubled. Both are long-horizon software engineering tasks, precisely the kind RL training targeted, and they line up with the training curves discussed below.

3.2 General and Agentic

General capability sits in the first tier, but unevenly. On "complete a professional task" benchmarks such as GDPVal, JobBench, and Automation Bench, the model is consistently top three. On hard terminal-operation tasks such as Terminal Bench 4.0, the gap is clear. Note that Terminal Bench 2.1 is near saturation (84 to 91 across the board) while 4.0 spreads the field, which makes 4.0 the more informative measure of current capability ceilings.

3.3 Visual

This is a Xiaomi in-house benchmark and the only visual evaluation reported, so cross-vendor comparability is limited. It does correspond to the frontend and design capabilities shown in the "Vibe World" demos.

3.4 Cyber

Cyber is the section that most needs a careful read:

• Tops CyberGym and MiMo Cyber Bench, but neither has GPT or Claude scores to compare against, so "leading" here means leading the open-source field.
• On the three benchmarks with closed-model baselines (SEC Bench Pro, ExploitGym, ExploitBench), Pro trails in every case, by 52 points to GPT 6 Astra on ExploitBench.
• Flash slightly edges Pro on CyberGym but drops sharply on exploitation tasks (25.3 on ExploitBench).

In short: strong on analysis and detection, weak on active exploitation. This pattern is consistent with the more conservative safety alignment typical of open models; the launch page does not comment on it.

---

4. Scaling RL: Results, Method, and Open Source

4.1 Training Results

The key result is out-of-distribution generalization. DeepSWE v1.1 was held out of training, yet scores rose steadily with RL steps:

Note: the official text quotes DeepSWE v1.1 as Pro 58.4→72.57 and Flash 48.8→65.68, which differ from the per-step table by rounding. AutomationBench and Visual Coding start at step 0.

Per-step scores and token consumption were published. Selected steps on DeepSWE v1.1 (all 30 steps are on the official page):

What the curves show:

• Tokens and scores rise together. Pro's per-task tokens grew from 138.7K to 214K (+54%); Flash's from 154.1K to 236.6K (+54%). Part of the gain comes from the model learning to explore and verify more, not only from a sharper policy.
• Flash consistently uses more tokens than Pro. On the same task, Flash spends 10% to 15% more tokens for a lower score, consistent with a smaller model compensating with more steps. It also means that under token-based billing, the real per-task cost gap between Flash and Pro is narrower than the list-price gap.
• Improvement is not monotonic. Pro regresses at steps 2, 6, and 13, reaches 72.6 at step 26, drops to 69.5 and 68.3 at steps 27 and 28, and returns to 72.6 at step 30. Keep this in mind when judging benchmark variance (no confidence intervals were published).
• A counterexample. On AutomationBench, Flash's tokens barely move (7.6K→9.3K) while its score rises from 45.2 to 52.7; on the same benchmark Pro's tokens double (6.9K→14.3K) for 46.5→55.0. Token growth is not a prerequisite for improvement, and the stated reward design that "steers the model toward shorter paths" does appear to take effect on some tasks.

4.2 Three Scaling Axes

The launch page groups RL compute scaling into three directions:

1. Larger batches, higher throughput: a fully asynchronous architecture, 1,568 samples per update, training at up to 1M context length, and 3.5 to 3.7B tokens per step.
2. More tasks, richer environments: a multi-task suite spanning coding, general agents, visual, and cyber, mixed across several harnesses so that gains in one capability reinforce the others.
3. More grader compute: relative comparison within each group gives long-horizon tasks more precise and diverse reward signals, closes a self-improvement loop, and steers the model toward shorter paths and fewer tokens per task.

4.3 Stability and Infrastructure

• As the run scaled up, the router was frozen to suppress training drift, which also indirectly confirms an MoE architecture.
• A defense against reward hacking was built spanning reward design, adversarial evaluation, anomaly detection, and cross-checking between verifiers.
• To support large-scale agentic RL across heterogeneous tasks: a unified trajectory representation and penalty mechanism, high-concurrency interaction across multiple agent frameworks, decoupling of the control plane from the data plane for large-scale trajectory migration, stabilized per-task sampling ratios within mixed batches, and consistency optimization between training and inference engines.

Some engineering reading between the lines:

• 1,568 samples and 3.5 to 3.7B tokens per step implies roughly 2.2M to 2.4M tokens per trajectory on average. Combined with 1M training context, these are multi-turn, long-horizon agent trajectories, not single-shot Q&A.
• "Relative comparison within each group" resembles GRPO-style methods: sample several trajectories per task and use within-group ranking rather than absolute scores as reward. This avoids training a value model per task and suits long-horizon tasks where absolute scoring is hard.
• Freezing the router is standard practice for MoE models in RL. It prevents expert assignment from drifting during policy updates at the cost of forgoing further routing-level optimization.
• Control/data plane decoupling and cross-harness trajectory migration indicate the training system schedules sampling from the same weights across multiple agent harnesses (different toolsets and execution environments). That is the main infrastructure difference between heterogeneous multi-task RL and traditional single-environment RL.

4.4 What Is Open-Sourced

The technical report, training environments, and RL code are all released so that researchers can reproduce and verify the results. This is what separates the release from most "open weights" launches: not only the outcome is open, but the process that produced it.

---

5. Demos: From Vibe Coding to Vibe World

Xiaomi positions V2.6 as extending natural-language programming from building software to building interactive worlds, combining 3D spatial reasoning, multimodal perception, and computer-use agent (CUA) capabilities. Demos cover: game development (multi-agent construction of 3D scenes with iteration on rendered output), 3D modeling in Blender, embodied simulation (multi-view camera input driving a Franka Panda arm in a closed loop), frontend and slide-deck generation (working with Figma and image/video generation tools), end-to-end video production (music, beat-synced editing, educational videos narrated by MiMo-V2.5-TTS), and music composition (orchestral and piano scores converted to MIDI by the model itself).

All of this is official demo material with no third-party benchmark. From a technical standpoint, the embodied simulation and game development demos are the most interesting because they require closed-loop correction from visual feedback over long horizons, which maps directly onto the long-horizon capability RL training targeted.

---

6. Research Case Studies

Xiaomi notes that neither case involved RL tailored to scientific research.

Co-scientist for materials research. Xiaomi's materials experts prompted Pro over several rounds to design a new metal-organic framework (MOF) for adsorbing per- and polyfluoroalkyl substances (PFAS, the "forever chemicals"). The model ran literature and patent searches, proposed hypotheses, assessed novelty, then invoked open-source computational tools to set up the simulation environment automatically, computed MOF–PFAS binding strengths (adsorption energy around -2.78 eV), and screened three UiO-67-type zirconium framework candidates for the wet-lab stage.

Formalizing mathematical proofs. Pro assisted researchers in completing a Lean 4 formalization of the full main theorem of Li and Yorke's Period Three Implies Chaos: for a continuous self-map of an interval, a period-three orbit implies periodic orbits of every positive integer period and an uncountable scrambled set. Guided by a researcher-designed exploration strategy and using subagent collaboration, the project grew to over 6,000 lines of Lean, fully verified by the Lean kernel with no unfinished placeholders. The model had received no Lean-specific post-training.

---

7. Availability

• Products: AI Studio, MiMo Code, and MiMo Desktop. MiMo Desktop leaves early access with its first official release, with Pro and Flash built in and UltraSpeed mode launched. The early-access program runs one more week; existing users should switch to the new model names.
• API: Xiaomi MiMo Open Platform, pricing unchanged from V2.5; Pro can be called in UltraSpeed mode.
• Third party: OpenRouter.
• Open source: Hugging Face weights, technical report, training environments, and RL code.

---

8. Assessment

Model selection

• Everyday agents, code completion, batch processing: choose Flash. Within 1 to 4 points of Pro on most benchmarks at one-third the price.
• Long-horizon complex tasks, research collaboration, anything requiring peak capability: choose Pro. On hard benchmarks such as GDPVal, Terminal Bench, SEC Bench Pro, and ExploitBench, the Pro–Flash gap widens to 6 to 22 points.
• Real-time chat and interactive creation: UltraSpeed, accepting the 10x unit price.
• Cyber workloads: usable for analysis and detection; do not expect it to approach GPT 6 Astra on active exploitation.

What stands out

• First open-source model to top the Artificial Analysis Intelligence Index, at unchanged prices, pushing the intelligence-versus-cost Pareto frontier outward.
• An unusually large generational jump: DeepSWE 19.0→71.9, Terminal Bench 4.0 1.5→34.9, Automation Bench 16.0→53.1.
• The live-streamed RL run and open-sourced environments and code make the "six days, thirty steps" claim reproducible by third parties rather than something to be taken from a chart.
• Continued improvement on a held-out benchmark is reasonably convincing evidence of RL generalization.

What to hold back on

• Top closed models retain clear leads on hard tasks: about 25 points behind GPT 6 Astra on Terminal Bench 4.0, 52 points on ExploitBench, and about 4 points behind Claude Fable 5 on OSWorld.
• Several "first place" results come from Xiaomi in-house benchmarks (MiMo Code Bench, Visual Coding, Cyber Bench) or benchmarks lacking closed-model baselines (CyberGym), limiting comparability.
• Vibe World demos and research cases are official showcases with no independent evaluation.
• The launch page does not disclose parameter count, active parameters, or serving context length (only that training ran at up to 1M context); these await the technical report.
• No confidence intervals or sample counts are given for benchmark scores. The RL curves show step-to-step swings of 3 to 4 points on DeepSWE v1.1, so leaderboard gaps under 2 points should not be over-interpreted.
• Reproduction cost: environments and code are open, but a single 30-step Pro run costs about $2.62M, which is out of reach for most labs. A more realistic verification path is to reproduce part of the Flash run or validate the generalization trend on the released environments.

Overall, MiMo-V2.6 is a release built around coding and agentic capability, with scaled RL as its methodology and full open-sourcing as its stance. Pro can now be compared head-to-head with Claude Opus 5 and GPT 6 Astra on aggregate capability, and Flash reproduces most of that at one-third the price. It raises the ceiling for open-source models by a clear margin while leaving the positions where closed models still lead plainly visible.

---

References

1. Introducing MiMo-V2.6 series — Xiaomi MiMo official launch page
2. Xiaomi MiMo Home — MiMo-V2.6-Pro
3. Artificial Analysis Intelligence Index v4.3, September 2026