Updates on Chinese AI models and their practical impact on API integration, model selection, and cost.
According to a MarkTechPost chronicle dated October 4, 2026, Alibaba's Qwen went from an invitation-only trial in April 2023 to a reported 2.4 trillion parameter open-weight model, with licensing shifting from free commercial use under 100 million monthly active users to Apache 2.0 across multiple sizes, alongside MoE sparsity, hybrid linear attention, and parallel open and closed product lines.
According to an October 4, 2026 report by MarkTechPost, Cantina Security, together with Yeta Labs, released the open-weight model apex-flash-1, fine-tuned on GLM-5.3-Flash with GRPO and released on Hugging Face under the MIT license, self-reporting 40 of 60 held-out vulnerability tasks solved and roughly 640GB of VRAM needed for BF16 weights.
According to an October 4, 2026 MarkTechPost report, Aleph Alpha released Kolibri-1, an English-German bilingual open-weight MoE model with 78.1 billion total parameters activating only 3.46 billion per token, supporting up to 1,048,576-token context and four inference-intensity levels, released on Hugging Face under Apache 2.0, with an FP8 checkpoint of about 78GB servable on a single B200 or H200 via vLLM.
According to an Ai2 blog post dated October 2, 2026, Ai2 has released AstaBrief 8B, post-trained from Qwen3-8B, which writes a research question and retrieved snippets into a cited report in a single pass and has gone live as the Fast mode of Asta's Generate Report feature; the blog says Fast mode averages 51.1 seconds per report, about 3.5 times faster than its Claude-driven Thinking mode, with weights and training data released as well.
According to a September 30, 2026 report by The Decoder, Anthropic's Frontier Red Team said Zhipu's open-weight model GLM-5.3 built 50 working exploits in 410 attempts on ExploitBench, approaching Mythos Preview's 56, and that its refusal safeguards can be stripped via abliteration; a separate US CAISI evaluation reached similar conclusions, saying it trails the best US models by about four months.
According to a September 30, 2026 Leiphone report, the author used Claude Code in the same environment to have same-day releases Qwen3.8-Flash and GLM-5.3-Flash build a personal workbench from scratch; both delivered runnable apps, with GLM leaning toward system and engineering completeness and Qwen toward personal workflow and prioritization design.
According to Machine Heart, RunningHub, under Haimayun, released its self-developed generative video super-resolution model RH Upscale, positioned to "upscale without changing content." The report says that in a 13-configuration comparison using 540P input and 1080P output, Ultra ranked first with a composite score of 0.539, and that five tiers are offered from Ultra to Low; Max reportedly takes 34 seconds for 2x super-resolution of a 5–10 second video, about 1/8.6 of Ultra's runtime.
According to an official NVIDIA blog post dated September 29, 2026, NVIDIA released Kumo Tabular, a tabular foundation model in three sizes from 28 million to 215 million parameters, published under the OpenMDW-1.1 license and hosted on Hugging Face. It requires no training or feature engineering, performing classification and regression in a single forward pass; the blog claims it ranks first on four benchmarks including TabArena.
According to Xinzhiyuan (新智元), an anonymous model called Space Bunny Alpha appeared on OpenRouter and OpenCode on September 23, and developers compared its tokenizer and speculated it came from MiniMax; four days later, on September 27, MiniMax launched M3.1-Flash-Preview, offering six adjustable reasoning tiers, a 1M-token context and text, image and video input. The report also documents test results covering bug fixes, web page generation, 3D and games.
According to a September 29, 2026 report by Xinzhiyuan, RunningHub launched H3 RH Enhanced, a video enhancement model deeply post-trained on the open MiniMax H3 base. It claims that across 2,000 test sets, multi-shot long-sequence narrative significantly outperforms Seedance 2.0, and the character-collapse rate in extreme long scenes drops by more than 60 percent. The report gives two cost figures: the body says 0.18 yuan per second, while the conclusion says 0.1 yuan per second, so the vendor's latest billing statement should be used to verify.
According to a September 28, 2026 report by Machine Heart, Dai Jifeng's team at Naive AI released its first model, Naive-N0.5-Flash, with 309B total parameters, 15.5B activated, native support for a one-million-token context, and a focus on coding and AI R&D. The report says the weights and inference code use the MIT license and that an API is planned, and it also discloses evaluation results including PaperBench and MLE-bench-30.
A study covered by a machine-intelligence outlet and published on arXiv in September 2026 reports that forward analysis and inference-time interventions on the public DeepSeek-V4-Flash-0731 weights found that mHC's four residual streams concentrate their read-write weights on about two streams, that the residual mixing matrices in layers 22 to 42 are close to the identity matrix, and that replacing deep-layer mixing with identity leaves the six-task average score almost unchanged.