On Beijing time 10 30, SenseTime open-sourced the multimodal model SenseNova Loop-9B-Preview, built on a Looped Transformer architecture: roughly 9B parameters, with 32 layers of weights shared and looped to perform 56 layers of computation per forward pass, officially claiming 27B-level performance along with 8 benchmark comparisons. Weights are on Hugging Face and code on GitHub under Apache 2.0. Community attention focuses on the choice of comparison targets as well as looped inference latency and KV cache overhead; independent testing and the technical report have yet to appear.
At the 2026 Global 1024 Developer Festival, iFLYTEK announced the device-cloud integrated Spark X2.6, the Xingchen TaaS platform, the SparkAudio 1.0 speech foundation model, and Agent OS. Spark X2.6 expands its context window to one million tokens and open-sources the 4B on-device model, with device-cloud routing accuracy improved to 96.1%; the iFLYTEK Open Platform developer base has reached 1187 ten thousand.
At the 2026 Tencent Global Digital Ecosystem Summit, Tencent announced that Hunyuan Hy4 is coming soon. Compared with the Hy4 preview, Hy4 reportedly delivers comprehensive, substantial improvements, with enhanced capabilities for long-horizon and real-world productivity tasks, improved chain-of-thought efficiency, and first-time visual understanding; the more practical Hy4 flash is also planned for near-term release. The article reviews the officially disclosed information and points still to be confirmed.
On 10 13, Modelbest released the open-source on-device multimodal large model MiniCPM-V 4.7, available in 35B-A3B and 1B versions. 35B-A3B adopts a MoE architecture with 350 billion total parameters and 30 billion activated per inference; according to the company, it requires only about 3 GB of memory to run, with a decoding speed of 15+ token/s. It supported Rockchip, NVIDIA, Intel, and Apple chip platforms as well as iOS, Android, and HarmonyOS from day one, with core optimizations from the proprietary LES lookahead expert scheduling technology.
According to the official news brief from DeepSeek, V4.1 Flash has been updated to version V4.1-Flash-1017, with the web version, APP, and API model simultaneously updated. According to the brief, the new model shows improved benchmark results while shortening task runtime and output length, lowering usage costs, and was optimized for platform frameworks such as DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy. Through self-feedback reinforcement learning training, the model has enhanced its project-building capabilities in areas such as game development and 3D modeling; DeepSeek API adds support for the Search and Fetch interfaces, which are now officially open.
According to BT Finance, Meta released the native multimodal open-source model Muse Spark Open on 2026/10/15, with 8160 billion total parameters and 420 billion active parameters. Built on a MoE architecture and deeply integrated with Apple's MLX framework, the model can reportedly run fully locally on two Mac Studios equipped with M5 Ultra and 512GB of memory. The article also notes that its performance data lacks independent third-party verification, and that on coding tasks it reportedly still lags some Chinese open-source models.
On 2026 17, 10, DeepSeek upgraded V4.1-Flash-1017 in place, with the API name and pricing unchanged, focusing on enhanced agent capabilities and wish-style programming. Third-party tests show that when all 10 coding tasks passed, spending was under 40% of Claude Opus 5.5, but took more than twice as long; cache gains depend on multi-turn session reuse, benchmarks are tied to evaluation frameworks, and hallucination correction, latency, and low-price competition remain open gaps.
The announcements section of Alibaba Cloud's official website published an upgrade notice for "Model Studio (Bailian)" regarding DeepSeek-V4.1-Flash, dated 2026-10-21, with the affected product listed as Model Studio (Bailian). The current announcement page does not provide details on the upgrade content, model behavior changes, or effective timing; verification should be based on Bailian platform documentation or subsequent announcements.
According to its official blog, Moonshot AI has released the first model in its K3.1 series, Kimi K3.1 Pro, along with the high-speed variant Kimi K3.1 Pro Highspeed. Both models support a 1000K-token context window, and the high-speed version runs at 2 to 3 times the speed of the standard version. Official data shows K3.1 Pro scored 65.70% on Terminal-Bench 4.0, with a notably lower hallucination rate than K3. The models are now live on Kimi.ai and other platforms as well as API, with full weights planned to be released before 4 11.
On 21, 10, 2026, OpenAI and Google Cloud announced a partnership under which GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna, and Codex agent workflows will enter Google Cloud Model Garden, with general availability expected within the coming weeks. Customers can purchase via Google Cloud Marketplace, with usage consolidated onto a single invoice and offsettable against committed spend; data is committed to staying within customer-designated boundaries and not used for training foundation models. Microsoft's exclusive distribution agreement expired in 4 2026, meaning OpenAI models now cover all three major US cloud providers.
Moonshot AI released Kimi K3.1 Pro in the early hours of 10/24, calling it the first model in the K3.1 series and its most capable 3-trillion-parameter model to date. It supports a 1000k context window and delivers significant improvements in coding, 3D creation, financial research, and long-horizon Agent capabilities. The model is now available via kimi.com, Kimi App, Kimi Work, Kimi Code, and Kimi API, with the full model weights planned for release before 11/4.
MiniMax announced on 10 (23) the open-sourcing of the MiniMax M3.1 Flash model weights, with 428B total parameters and 23B activated parameters. It delivers comprehensive capability improvements over the previous-generation M3 model, with inference efficiency enhanced through model-wide sparse attention, quantization-aware training (QAT), and Power-aware scheduling. After launching on 10 (19), the model processed over 100 trillion tokens from real users within 72 hours, and secured day-one adaptation from multiple domestic and international chip vendors and inference cloud platforms at open-source release.