On 10 23, MiniMax open-sourced its multimodal large model MiniMax M3.1 Flash. On the same day, Moore Threads' flagship AI training-and-inference compute card MTT S5000 and its proprietary MUSA software stack completed Day-0 adaptation of the model. With the adaptation timing synchronized with the model's open-source release, the model can run on Moore Threads' domestic GPU platform as soon as it is open-sourced, though the source provides no details on performance, deployment scale, or ecosystem.
Ling-3.1-flash, open-sourced by Ant Group's inclusionAI, is a hybrid-architecture MoE language model with 560B total parameters and 25B activated per token, with FP8 quantized weights. According to a CSDN blog compilation, as of 2026 10, 15, the SGLang official deployment page verified two recipes: 8×H200 and 4×GB300; the official 54 safetensors shards total approximately 573.8 GB. Deployment requires multi-GPU resources and a dedicated image; the context length has three definitions — 1M nominal, 256K on the official service, and 262,144 in the weight config — and first deployments should start with a short context.
On 20 10, DeepSeek announced it would fully open-source its internal inference framework DeepInfer under the Apache-2.0 license, along with the MIT-licensed token encode/decode component deepseek-recipe. DeepInfer has already supported DeepSeek V4.1-Flash large-scale production inference and currently supports NVIDIA SM100 and Huawei Ascend 950 platforms, but it is an offline inference engine without online serving or a standalone KV cache system, and supports only the DeepSeek series of models.
On 15 10, 2026, StepFun open-sourced the Step 5 Preview model weights, available in both FP8 and BF16 formats. The model uses a sparse MoE architecture with 600B total parameters and 27B activated parameters, contains 92 Transformer blocks, and supports up to 1M tokens of context; the company says it targets long-horizon Agent tasks in software engineering and professional knowledge work. Commercial use requires reviewing the additional conditions of the StepFun Model License 1.0.
MiniMax officially released MiniMax M3.1 Flash. Officially, it previously ran under the anonymous model name Space Bunny, cumulatively processing over 200 trillion tokens. The new model focuses on agentic tasks, coding, Computer Use, and cybersecurity, with officially published benchmark results including AutomationBench-AA, BankerToolBench 76.2, DeepSWE v1.1 73.7%, OSWorld V2 61.7, ALE 29.6, and CyberGym 82%. Officials claim nearly 3x faster online inference speed and a 70% cost reduction versus the previous generation, priced at input $0.15, output $0.60, cached input $0.015 (per 1M tokens). The model is now available via MiniMax API, M Plan, and MiniMax Code, with weights to follow.
On 10 14, 2026, Qianwen Office CEO Chen Yusen disclosed at the e-commerce edition launch event that more than two months after launch, the product's paid user share in the e-commerce sector exceeded 20%. A merchant case presented at the event showed that after adopting the AI agent, new store preparation time was compressed from one month to 7 days, initial investment dropped from over 1 thousand yuan to a few hundred yuan in Token costs, and daily operations work takes only 2 to 3 hours.
According to Meta's official press release, Meta launched two models, Muse Wonder and Muse Wonder Full, on 10/13, describing them as the company's most capable models to date, achieving industry-leading performance on personal and professional agent tasks. The new models strengthen professional-domain knowledge in law, finance, and healthcare as well as cutting-edge coding abilities, and introduce two new features, Active Memory and Active Focus. They are available starting today via Meta Model API and Muse Code, with subscribers getting first access.