MiMo-V2.6-Flash is Xiaomi's open-weights efficiency tier: a 309B-total / 15B-active sparse Mixture-of-Experts model with a 1M-token context window, hybrid sliding-window + global attention, and multimodal (image) input. Positioned as a low-cost agentic/coding workhorse at $0.14/$0.28 per 1M tokens, it trades raw frontier intelligence for cost efficiency.
MiMo-V2.6-Flash (OpenRouter slug: xiaomi/mimo-v2.6-flash, Benchmark Heaven ID: mimo-v2.6-flash::default) is one of two flagship-adjacent tiers released by Xiaomi in the MiMo-V2.6 series on September 21-22, 2026, alongside the larger MiMo-V2.6-Pro sibling. It is an open-weight (MIT-licensed) sparse Mixture-of-Experts foundation model explicitly engineered for the efficiency/cost tier rather than peak reasoning.
Architecture: 309B total parameters with only 15B activated per token, using 256 routed experts with 8 activated per token. The transformer stack comprises 48 layers: 39 sliding-window attention (SWA) layers with a 128-token window (64 Q heads / 8 KV heads) interleaved with 9 global-attention layers (64 Q heads / 4 KV heads). Head dimensions are 192 (QK) and 128 (V). Hidden size is 4,096. This hybrid attention design is what enables native 1M-token context at manageable compute. It supports text-plus-image input (the model card additionally lists video and audio modalities). Training followed Xiaomi's RL-scaling program (fully asynchronous RL, large rollout batches, sequences up to 1M tokens), and the model is distributed with a multi-token-prediction / speculative-decoding head (5 SWA layers, 1,024-token window) — though backend support for that path varies by host.
Positioning vs. the October 2026 frontier: MiMo-V2.6-Flash is a value-oriented open-weights option. Its Artificial-Analysis-calibrated intelligence index of 37.9 and Vals Index of 59.6% place it below frontier closed models (GPT-6 Astra, Claude Opus 5, Gemini 3.8 Pro) and below its own 1T+ MiMo-V2.6-Pro sibling, but at a fraction of their price and with a very large 1.05M context window. It is best understood as a high-throughput, low-cost agentic/coding/automation tier rather than a reasoning-max model.
Throughput: Benchmark Heaven records an empirical ~51 tokens/sec; third-party OpenRouter telemetry shows a median throughput that has varied (roughly 19-30 tps across snapshots) with ~4.9s latency on the current stable page. The wide spread reflects differing backends, prompt/generation lengths, and reasoning-token overhead rather than a single hardware spec.
Data-integrity note: only benchmark scores explicitly evaluated on MiMo-V2.6-Flash are reported. Scores for the larger MiMo-V2.6-Pro sibling, the MiMo-V2.6-Distill-Qwen-9B, and prior MiMo generations were deliberately excluded. The benchmarks below are ADDITIONS beyond the verified Benchmark Heaven ground-truth baseline, which remains unaltered.
309B total parameters / ~15B activated per token; 256 routed experts with 8 activated per token. 48 transformer layers: 39 sliding-window-attention layers (128-token window; 64 Q / 8 KV heads) + 9 global-attention layers (64 Q / 4 KV heads); QK head dim 192, V head dim 128; hidden size 4,096. Native 1.05M-token context. Multimodal input (image; model card also lists video/audio). Trained via Xiaomi's asynchronous RL-scaling program; ships with an MTP/speculative-decoding head (5 SWA layers, 1,024-token window). Open-weight, MIT license. Released September 21-22, 2026.
Recommended Workloads & Primary Use Cases
Long-context document, image, and mixed-media analysis up to ~1M tokens at very low cost
High-volume agentic automation and tool-use pipelines where per-token cost dominates
Code generation, repo-level refactoring, and code migration tasks
•Reasoning-token behavior is not officially documented for the public API; do not assume a visible think/reasoning field is returned, and account for hidden-token overhead when budgeting latency.
•Effective throughput is highly backend-dependent: Benchmark Heaven measures ~51 tps but OpenRouter snapshots have reported ~19-30 tps with ~4.9s latency, so TTFT and sustained TPS should be measured per provider.
•The 1.05M context window is nominal; effective long-range recall degrades and no variant-specific MRCR v2 long-context score is published, so validate retrieval/recall at target context depths before relying on full-window prompts.
•Several reported scores (DeepSWE v1.1, ProgramBench, MiMo Code Bench, AutomationBench, Toolathlon-Verified, Agents' Last Exam) originate from Xiaomi's own model-card comparison table and are vendor self-reported, not independently replicated.
•As an efficiency tier, it trails its larger MiMo-V2.6-Pro sibling and frontier closed models on reasoning-heavy and legal/research benchmarks (e.g., Harvey LAB-AA 11.3%), so avoid it for tasks demanding peak multi-step reasoning accuracy.
ARMES Zero Data Retention
Calls to MiMo-V2.6-Flash are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by Xiaomi.
