MiMo-V2.6-Flash-RL

MiMo-V2.6-Flash-RL is the efficiency-balanced checkpoint of Xiaomi MiMo's V2.6 series: a 310.76B-parameter sparse mixture-of-experts model with 15B active parameters per token, a 1M-token context window, and native handling of text, image, video, and audio. Weights ship on Hugging Face under MIT, no gated access, no custom license, no revenue caps. That combination is unusual at this scale and is the main reason this model matters to anyone running inference on their own hardware.

The "RL" suffix is load-bearing. This checkpoint comes out of a single mixed reinforcement learning run, coding, general agents, visual tasks, and cybersecurity trained in the same batch rather than in separate per-domain passes. Xiaomi's bet is that mixing environments and graders in one run lets capabilities reinforce each other and transfer to agent harnesses the model never saw during training. They also scaled the reward signal itself: binary pass/fail can't rank two passing solutions, so groupwise agentic grading builds task-specific rubrics from contrasting rollouts and redistributes advantage toward shorter, more token-efficient trajectories. Practically, that means you should expect tighter outputs and better tool discipline than a pure SFT or per-domain RL checkpoint at the same size.

Position it against the field. It sits in the same weight class as Qwen3-235B-A22B and DeepSeek-V3 (671B total / 37B active), but with a materially longer context window, native audio and video input, and an MIT license instead of a bespoke community license. On the MiMo API platform it's listed at $0.14 per million input tokens and $0.28 per million output, the Flash SKU is priced as the high-frequency workhorse, not the flagship.

Three numbers matter for anyone planning hardware:

• 310.76B total parametersthe full weight set you must store and load.
• 15B active parametersroughly what's read per token during decode.
• ~4.8% activation ratiothis is what makes a 310B model fast enough to be practical.

That gap between total and active is the entire value proposition. Compute cost per token tracks the 15B active slice, so FLOPs are comparable to a dense 15B model. Memory footprint tracks the 310.76B total, so VRAM is not. You cannot stream experts off an NVMe drive fast enough for interactive use, plan on holding the full quantized weight set resident in VRAM or unified memory.

Supporting components include a 681M-parameter MiMo ViT for vision (28 layers, 24 sliding-window attention + 4 full attention), a 308M AudioTokenizer paired with a 127M audio patch encoder, and a 5-layer multi-token prediction (MTP) speculative decoder that improves decode throughput without changing output quality.

The 1M-token context window is the other headline spec. It's enough for whole-repository code analysis, long tool-call traces from multi-session agent runs, hour-scale audio, and multi-minute video with frame-level reasoning. Note that KV cache at 1M tokens is itself a serious memory cost, budget for it separately from weights, and expect to use context caching to keep it manageable.

Coding and agentic work. The mixed RL run included coding and general agent tasks with harness-level training, so this checkpoint is built for tools like MiMo Code, OpenCode, Cline, and Claude Code-style harnesses. Realistic workloads: long-context refactors across a full repo, multi-step bug triage with file reads and test runs, PR review against a full codebase in one context window. Function-calling is native, not bolted on.

Vision and video. Native image input handles UI screenshots, document layouts, charts, and diagrams. Video understanding covers temporal reasoning rather than just frame captioning, useful for screen-recording QA, monitoring footage, or demo walkthroughs. The 1M context lets you keep an entire video's frame tokens and transcript in one window.

Long-form reasoning. The MTP decoder and RL-tuned trajectory steering favor shorter chains over verbose ones. For math, logic, and structured analysis, expect it to be competitive within the Flash tier rather than at Pro-level, this is the efficiency checkpoint, not the flagship.

Cybersecurity. Explicitly part of the RL training mix. Realistic uses: log triage, vuln triage on diffs, config review. Not a substitute for tooling, but a strong reasoning layer on top of it.

Languages. English and Chinese are the declared targets on the model card. Other languages will work but are not the training focus.

Be honest with yourself about the hardware math before downloading. At 310.76B parameters, weight sizes are roughly: FP16 ~620 GB, Q8 ~320 GB, Q4_K_M ~170–185 GB. The active-parameter advantage does not reduce storage or VRAM requirements.

What cannot run it: a single RTX 4090, 5090, or any 24–32 GB consumer GPU. Even a 128 GB M4 Max is not practical, Q2 quantization lands near 100 GB before KV cache, and quality loss is severe.

Realistic hardware for local inference:

• Mac Studio M3 Ultra 512 GBthe best single-box option. Q4_K_M at full context, expect roughly 25–45 tokens/sec depending on batch size and context length.
• Mac Studio M2 Ultra 192 GBQ4_K_M fits tightly. Short context only; long-context work will thrash.
• 4× RTX 6000 Ada (48 GB each, 192 GB total)comfortable Q4_K_M with tensor parallelism in vLLM or SGLang.
• 8× RTX 4090 (24 GB each, 192 GB total)workable Q4_K_M, but PCIe interconnect throttles throughput. Expect 15–30 tokens/sec, worse with large batches.
• 2× RTX PRO 6000 Blackwell (96 GB each)Q4_K_M with KV-cache headroom; the densest option per rack unit.

Quantization guidance: Q4_K_M is the default for most people, best quality-per-byte, and it's the size every hardware estimate above assumes. Move to Q5_K_M or Q6_K if you have 30%+ memory headroom and care about subtle reasoning degradation. Avoid Q2 and Q3 unless you're testing on constrained unified-memory Macs; the MoE routing already loses expert capacity at low bit widths.

Quickest start: Ollama if it publishes a MiMo-V2.6-Flash-RL tag, or llama.cpp/vLLM directly from the Hugging Face weights. For multi-GPU, vLLM with tensor parallelism is the production path.

vs. Qwen3-235B-A22B. Roughly 75B fewer total parameters means Qwen3 is meaningfully easier to run, a 192 GB box holds it at higher quantization with more KV headroom. Qwen3 has no native audio or video input and a 256K-class context in most deployments. Choose Qwen3 if you need maximum quality-per-GB on text-only work. Choose MiMo-V2.6-Flash-RL if you need the 1M window, native multimodal input, or the RL-tuned agent behavior.

vs. DeepSeek-V3 (671B / 37B active). DeepSeek-V3 is a stronger raw reasoner on many benchmarks, but the weight set is more than double. On the same 192 GB hardware, MiMo-V2.6-Flash-RL runs at Q4_K_M while DeepSeek-V3 needs aggressive quantization that eats into its advantage. If you have 8× H100-class accelerators, DeepSeek-V3 wins on quality. If you're on a Mac Studio or a 4–8 GPU workstation, MiMo is the practical choice.

The honest summary: pick MiMo-V2.6-Flash-RL when long context, multimodal input, permissive licensing, and agent-harness behavior matter more than squeezing the last few benchmark points out of a text-only model.