MiMo V2.6 Pro and Flash launched together: how strong are they?

MiMo V2.6 is not a Pro-only release. Xiaomi published Pro-RL and Flash-RL together on 21 September 2026. Pro activates about 42B parameters per token from a 1.02T total, while Flash activates about 15B from 309B. Both are sparse MoE systems that keep all experts available but route each token through only part of the network. Reading Flash's 15B active count as the file size of an ordinary dense 15B model would severely understate its requirements.

The architecture is closely related. Pro has 70 layers and Flash 48, interleaving global attention among many short-window attention layers. Pro carries 384 routed experts and Flash 256, with eight selected per token. Both add a five-layer MTP decoder so generation need not predict only one future token at a time. Flash is therefore less a stripped-down trial edition than the same design scaled for a smaller serving footprint.

The official results stand out most on tasks that use tools rather than merely recall stored knowledge. DeepSWE v1.1 scores 71.9 for Pro and 67.9 for Flash; ProgramBench is 26.5 and 26.0. AutomationBench is 53.1 and 52.3, while Toolathlon is 76.9 and 73.6. Pro leads, but Flash does not fall behind in proportion to its much smaller scale. The results support the release's emphasis on editing repositories and operating tools.

The gap is also narrow in computer-use tests. OSWorld Verified is 82.0 for Pro and 80.8 for Flash, and JobBench is 62.0 versus 61.2. Terminal Bench 2.1 is 89.9 versus 87.6, and Visual Coding 72.3 versus 71.5. Terminal Bench 4.0 is different, widening to 34.9 versus 28.8. Flash is not equal to Pro everywhere, but it is close often enough to be a serious deployment option when cost and throughput matter.

Flash has roughly 30% of Pro's total weights and about 36% of its active compute, yet the reported scores do not shrink in the same proportion. Both models share the same input modes, long context and tool-use training approach, while Flash retains a large expert pool and the MTP decoder. Flash even scores 95.1 on CyberGym against Pro's 94.0, and trails by only 0.5 on ProgramBench. The larger model does not win every task automatically.

The word Flash is not a latency guarantee. Real speed depends on weight precision, tensor and expert parallelism, interconnect bandwidth, MTP acceptance and concurrent load. The official card demonstrates Flash with an SGLang TP8/DP2 setup and a vLLM TP4 setup. Those examples describe stable multi-accelerator serving rather than a peak result on one consumer GPU.

Both models accept image, video and audio as well as text. They include a roughly 681M vision encoder plus an audio tokenizer and patch encoder, with a native context of 1,048,576 tokens. The design targets workflows that inspect a codebase, read screenshots and process video or audio in one model. Multimodality is part of the base architecture rather than a separate accessory.

A 1M context is not a default setting to enable on day one. Long inputs expand KV cache and prefill time, and visual or audio tokens increase the cost of each request. A real deployment starts from the document lengths and concurrency it actually needs, then budgets the cache. Supporting a 1M context and serving it repeatedly at low latency are different hardware claims.

The files in the official repositories total about 177.7 GB for Flash weights and about 573.5 GB for Pro. KV cache and runtime workspace come on top. These models do not place only the active parameters on one 32 or 48 GB GPU and discard the rest. Even Flash should be evaluated as a multi-accelerator or very-high-memory server workload.

The official SGLang example for Pro uses two nodes with TP16 and EP16, while its vLLM example uses TP8. Even Flash starts at TP4 in the official examples. The open weights can be operated without an outside API, so this is local AI in the ownership sense. In practice, local often means an internal multi-GPU server rather than one desktop. Open weight and personal-computer model are not synonyms.

Pro is the reference when a research setup serves only a few users and maximum task performance matters most. Flash is the practical starting point when many coding-agent and automation requests must be served and the relevant benchmarks show only a small gap. It leaves more room for throughput or a smaller first deployment. If the workload resembles a test where the gap is large, such as Terminal Bench 4.0, Pro still needs direct evaluation.

Neither model is a sensible reason to buy hardware before testing. Build an evaluation set from repository edits, browser actions and security-analysis tasks that currently fail, then compare the models on rented capacity or an existing cluster. Published scores narrow the shortlist. The final choice should follow measurements of success rate, per-request latency, concurrency and operating cost on your own work.