AiCybr

Xiaomi MiMo-V2.6 Pro: 1.02T MoE, 1M Context and MIT-Licensed Weights

By AiCybr Editorial · Published September 22, 2026

Xiaomi has released MiMo-V2.6, a new family of open-weight multimodal models led by MiMo-V2.6 Pro. The flagship uses a sparse mixture-of-experts architecture with 1.02 trillion total parameters and 42 billion activated per token, supports text, image, video and audio input, and provides a 1 million-token context window. Xiaomi has published the Pro checkpoint under the MIT license.

The release pairs unusually large model capacity with a comparatively small active parameter count. MiMo-V2.6 Pro has 70 transformer layers, 384 routed experts with eight activated per token, and a five-layer speculative decoder. Its multimodal stack includes a 681M-parameter vision encoder plus dedicated audio encoders. Output is text.

For developers evaluating hosted access, Xiaomi lists MiMo-V2.6 Pro at $0.435 per million uncached input tokens and $0.87 per million output tokens. Cache-hit input is $0.0036 per million tokens. The hosted model supports up to 128K output tokens alongside its 1M-token context window.

Pro and Flash target different deployment tiers

MiMo-V2.6 includes Pro and Flash as the two main native multimodal models. The published Flash checkpoint uses a smaller architecture aimed at lower-cost serving, while Pro is the flagship for long-horizon coding, agent, computer-use and research workloads.

The Pro architecture uses 60 sliding-window-attention layers and 10 global-attention layers. Xiaomi's model card lists 128 query heads and eight key/value heads, a hidden size of 6,144, and a sliding window of 128 tokens. The combination is designed to reduce long-context KV-cache pressure while retaining periodic global attention.

The 42B active-parameter figure describes compute routing per token. Self-hosting capacity planning still has to account for the full checkpoint representation, quantization format, KV cache and multimodal components. This places the flagship in multi-accelerator or high-memory server territory for practical local deployment.

Independent benchmark puts Pro at the top of the open-weight group

Artificial Analysis independently measured MiMo-V2.6 Pro at 46 on its Intelligence Index, placing it at the top of the open-weight models in the current comparison. Independent launch coverage also reports roughly 130 output tokens per second through Xiaomi's hosted API, although serving speed varies with provider configuration, workload and concurrency.

Xiaomi publishes a broader set of agent evaluations in the model card. Its reported Pro results include 71.9 on DeepSWE v1.1, 89.9 on Terminal Bench 2.1, 82.0 on OSWorld-Verified, 94.0 on CyberGym and 66.3 on SEC Bench Pro. These are Xiaomi-run measurements tied to the listed harnesses and evaluation settings; they provide workload-specific results alongside the independently measured composite index.

The technical report says Xiaomi scaled reinforcement learning across coding, general-agent, visual and cybersecurity environments. It describes asynchronous GRPO training with large rollout batches and a groupwise grading system intended to distinguish quality among trajectories that already pass binary task checks.

MIT licensing broadens deployment options

The exact MiMo-V2.6 Pro checkpoint on Hugging Face is tagged MIT, giving developers broad rights to use, modify and redistribute the released weights subject to the license terms. Xiaomi has also published the technical report with the checkpoint, making the architecture and training approach directly inspectable.

That combination makes MiMo-V2.6 Pro relevant beyond its launch benchmark position. Teams can evaluate the same released artifact through hosted inference or self-managed infrastructure, while the 1M-token context and multimodal inputs make repository-scale coding, long agent traces and mixed-media analysis the most obvious evaluation targets.

Deployment fit

MiMo-V2.6 Pro is best evaluated as a server-class open-weight model. Its sparse routing lowers active compute relative to a dense trillion-parameter model, while the full checkpoint remains large enough that ordinary consumer GPUs are outside the straightforward deployment path. Flash is the more practical member of the family when serving cost and hardware capacity dominate the decision.

For API users, the combination of the independently measured Intelligence Index result and Xiaomi's low published token pricing makes Pro particularly interesting for agentic workloads where repeated tool calls can amplify inference cost. For self-hosters, memory capacity, quantization support and inference-engine compatibility should be validated against the exact checkpoint before hardware planning.

Sources

Continue reading

Artificial Intelligence MiniCPM5-2B: 2.52B Parameters, 128K Context and Apache-2.0 Local AI

OpenBMB MiniCPM5-2B is a 2.52-billion-parameter dense model with 131,072-token context, Apache-2.0 weights and broad local inference support. Here is what its benchmarks and deployment options show.

Artificial Intelligence Abacus.AI Smaug Models: Agentic Fine-Tunes of Kimi K3, DeepSeek V4 Flash and Qwen3.8

Abacus.AI refreshed its Smaug open-weight model line for agentic workloads with Smaug Agentic on Kimi K3, Smaug Flash on DeepSeek V4 Flash and Smaug Mini on Qwen3.8 27B. This guide covers architecture, context, benchmarks, deployment and licensing details.

Artificial Intelligence K2 Horizon: Six Fully Open AI Models from 0.9B to 375B, with Training Data and Code

IFM released six K2 Horizon models from 0.9B to 375B parameters under Apache 2.0, alongside training code, checkpoints, logs, data or reconstruction recipes, quantization support and deployment tooling.