MiMo-V2.6-Pro vs GLM-5.2: Two Open-Weight MoEs, and the Benchmark You Must Not Subtract

The laziest version of this comparison puts Xiaomi MiMo-V2.6-Pro's 72.57 next to GLM-5.2's 46.2, calls it a 26-point win, and publishes. It is wrong for a reason that has nothing to do with which model is better: those two numbers are not the same measurement. Xiaomi's 72.57 is average-of-three on its own harness with mini-swe-agent and has never been submitted to a public board; the 46.2 is a Pass@1 figure from a different harness entirely. The honest comparison between Xiaomi MiMo-V2.6-Pro and GLM-5.2 is a different one, on a ruler both were actually run against — and on that ruler, the gap is real but it is twelve points, not twenty-six.

On the Artificial Analysis Intelligence Index v4.3.2, read on 22 September 2026, Xiaomi MiMo-V2.6-Pro scores 46. GLM-5.2 scores 34. Both are open-weights mixture-of-experts models from Chinese labs, both run a 1M-token context window, and both are priced cheaply enough that the choice between them is about capability and serving, not budget. That is where the resemblance ends, because GLM-5.2 has been superseded by its own maker, and the page this comparison is written on carries a deprecation notice.

What each model is, and what state it is in

GLM-5.2 is Z.ai's flagship, released 16 June 2026. It is a mixture-of-experts transformer at roughly 753 billion total parameters with about 40 billion active per token, and it was reported to be trained entirely on Huawei Ascend accelerators rather than Nvidia hardware. It ships open weights under the MIT licence with FP8 and INT4 quantisations, holds 1M tokens of context with up to 131,072 tokens of output, and lists at $1.40 per million input tokens and $4.40 per million output on Z.ai's own API, with cached input at $0.26 and a blended rate of $0.90. Artificial Analysis measured it at 72.6 output tokens per second and 5.98 seconds to first token, at a cost of $1,559.05 to run the full index.

Xiaomi MiMo-V2.6-Pro went generally available on 22 September 2026, with its reinforcement-learning checkpoint repositories appearing the day before. It is a sparse mixture-of-experts model at 1.02 trillion total parameters and 42 billion activated per token — a larger total than GLM-5.2 and a near-identical active count — with a 1M-token context window, native multimodality across text, image, video and audio, and MIT-licensed weights. Xiaomi kept the previous generation's pricing rather than repricing upward, so it lists at $0.43 and $0.87 per million tokens with a 99% cache discount and a blended $0.18.

• Weights — both MIT-licensed and downloadable; this is the only category where the two are genuinely equal

• Active parameters — MiMo-V2.6-Pro 42B per token; GLM-5.2 about 40B. Near-identical, which makes the score gap a training result rather than a scale result

• Total parameters — MiMo-V2.6-Pro 1.02T; GLM-5.2 roughly 753B

• Index score — MiMo-V2.6-Pro 46; GLM-5.2 34, both on Artificial Analysis v4.3.2

• List price — MiMo-V2.6-Pro $0.43 / $0.87 per 1M; GLM-5.2 $1.40 / $4.40, blended $0.18 against $0.90

• Cost to run the index — MiMo-V2.6-Pro $206.66; GLM-5.2 $1,559.05

• Speed — MiMo-V2.6-Pro 134.3 output tokens/second; GLM-5.2 72.6

• Time to first token — MiMo-V2.6-Pro 2.15 seconds; GLM-5.2 5.98 seconds

• Status — MiMo-V2.6-Pro current and GA; GLM-5.2 deprecated on Artificial Analysis, which now benchmarks it only at the default 10k input workload

The deprecation notice is the headline

Z.ai has since shipped GLM-5.3, and the Artificial Analysis page for GLM-5.2 says so directly, marking the model deprecated and limiting continuing benchmark work to a single default workload. That changes what this page is for. If you are choosing a new model today, the matchup you actually care about is MiMo-V2.6-Pro against GLM-5.3, not GLM-5.2 — and on the same v4.3.2 index, GLM-5.3 at maximum effort scores 45 against MiMo-V2.6-Pro's 46. One point. That is a real fight, and it is a completely different article.

The GLM-5.2 comparison is still worth making, for one specific reason: it is the cheapest possible illustration of how fast the open-weights frontier moved between June and September 2026. Twelve index points in about fourteen weeks, with the active parameter count essentially unchanged. Whatever produced that gain, it was not scale.

What GLM-5.2 was good at, and whether it still is

GLM-5.2's launch material led on coding and long-running agentic work, and those numbers still hold up as a description of the model: SWE-bench Pro at 62.1 against 58.4 for GLM-5.1, Terminal-Bench 2.1 at 81.0 against 63.5, and FrontierSWE at 74.4 against 30.5. On the knowledge and mathematics side it reported GPQA-Diamond at 91.2, AIME 2026 at 99.2 and Humanity's Last Exam at 40.5. It led all models on one long-horizon agentic coding benchmark at the time and ranked second on a public web-development leaderboard. Those are vendor-reported figures on the vendor's own harnesses, and the usual caveat applies to every one of them.

Xiaomi's own published numbers for MiMo-V2.6-Pro are not comparable to them, and it is worth being precise about why rather than gesturing at it. Xiaomi reports the model moving from 58.4 to 72.57 on DeepSWE v1.1 across its reinforcement-learning run, evaluated with mini-swe-agent at average-of-three, on its own grader, with the run documented as 30 steps and roughly 750,000 trajectories completed in under six days at a published spend of about $2.62 million for the Pro model. None of that is a leaderboard entry. A leaderboard entry is a claim about a specific configuration under stated conditions that a third party can re-run, and Xiaomi's is not one. A separate tracker lists GLM-5.2 at 46.2 on DeepSWE — Pass@1, a different metric on the same task family — and putting 72.57 beside 46.2 to produce a 26-point margin is arithmetic on two different scales. It should not be done, and it is the most common error in circulation about this pair.

Deciding between them, and the routing that makes it cheap

If both are open weights under the same licence, the deciding factors are what you can run and where you can call it. GLM-5.2 is on OrcaRouter as z-ai/glm-5.2 — one OpenAI-compatible endpoint, provider list price passed through with 0% markup, so a Z.ai price change is live on our side the same day. Xiaomi MiMo-V2.6-Pro is not on OrcaRouter; no Xiaomi model is, and reaching it means Xiaomi's own platform or one of the third-party catalogues carrying it. We would rather state that than let a model page imply otherwise.

The useful consequence is that you can hold the incumbents on one key and evaluate the newcomer against them without a second contract. If MiMo-V2.6-Pro wins on your prompts, you have learned that cheaply. If it does not — and the twelve-point index gap is smaller than the per-token price gap is large, so the outcome genuinely depends on your workload — you have lost nothing but the test. Routing the two behind one endpoint with automatic failover also answers the practical objection to a model that launched this week: a single serving route is a single point of failure, and a lane you can fall back from is what makes a new model safe to put in front of real traffic.

The short answer

If you are on GLM-5.2 today and it is working, the upgrade question is not MiMo-V2.6-Pro — it is GLM-5.3, on the same lineage, with a one-point gap to the Xiaomi model and none of the migration cost. If you are picking a lane from scratch and want open weights with the best measured score of the three, MiMo-V2.6-Pro's 46 is the number, and its 134.3 tokens per second and $0.13 per index task are the reasons it is not a narrow win. If you want the safest of the three, GLM-5.2 is the wrong answer for a reason that has nothing to do with how good it was in June: it is the only one of the three that its own vendor has stopped developing.

OrcaRouter puts 200+ models behind one OpenAI-compatible endpoint at provider list price with 0% markup, so a vendor price change is live the same day.