Updates on Chinese AI models and their practical impact on API integration, model selection, and cost.
According to an October 21, 2026 MarkTechPost report, Recursive released its first model, Flint, a 40B-total-parameter MoE coding model with roughly 2.5B active parameters, published under Apache 2.0 weights at about 23GB in 4-bit, runnable on a 48GB Mac or GPU.
According to an official IBM blog post published on October 21, 2026, the Granite Speech series has three updates: the new model granite-speech-5.0-2b supports six-language transcription and bidirectional translation with English; the Turbo CTC 470M version adds multilingual support and improves RTFx by about 1.6x; and 4.1-2b-plus improves speaker attribution and timestamps. All three are released under the Apache 2.0 license.
According to an InfoQ report on October 20, 2026, Jiuzhang Cloud released Alaya-DFlash2, a speculative decoding acceleration scheme for GLM-5.3 that includes a parallel-drafting draft model and a path selection mechanism, with weights published. The vendor's own evaluation claims single-concurrency decode speed at 8xH200, TP=8 is 3.41x that of native autoregressive decoding, and that accepted lengths across 1k to 32k low-entropy contexts are higher than two comparison schemes. These numbers and comparisons are vendor self-reported benchmarks and await third-party reproduction.
According to a report by The Decoder on October 19, 2026, MiniMax released the reasoning model M3.1 Flash; Artificial Analysis gives it an intelligence index score of 42, 13 points above the previous M3, at about $0.22 per task, with a 1 million token context and text and image input, and a planned open-weight release on October 23.
The Decoder reported on October 19, 2026 that Kling AI released 4.0: up to 30 seconds per generation, up to ten keyframes, 8,000-token prompts, new stereo audio and lip sync in more than a dozen languages, plus a Flash version capped at 720p and 20 seconds; API pricing and support status are still listed as coming soon.
According to a Xinzhiyuan report dated October 19, 2026, MiniMax released M3.1 Flash, which is live on MiniMax API, M Plan and MiniMax Code, with the official CLI's default text model switched. The report says it ranks at or near the top on Automation-Bench and LOCA-Bench 256K, among others, and comes with 1M-token context and five effort levels.
According to a MarkTechPost report dated October 19, 2026, MiniMax has launched M3.1 Flash on its API: 428B total parameters, roughly 23B activated, 1M token context, priced at $0.15 input and $0.60 output per million tokens. MiniMax said weights are expected on October 23, with the license not yet disclosed.
According to an InfoQ report on October 18, 2026, Meta opened the Muse Spark Open weights on October 15: a sparse MoE with 816 billion total parameters and roughly 42 billion active parameters, supporting a context of about 1.05 million tokens and text, image, audio, and video input, deployable on two M5 Ultra Mac Studios via MLX. The report says its Terminal-Bench 2.1 and DeepSWE 1.1 scores trail GLM 5.3 max and Kimi K3 max, and that it is licensed under the Muse Model License.
According to an October 17, 2026 report by THE DECODER, DeepSeek released V4.1-Flash-1017, which the company says outperforms larger rivals on coding and agent tasks at lower per-task cost, with weights published on Hugging Face and ModelScope and its in-house search and scraping APIs opened to developers for the first time.
According to a Xinzhiyuan report on October 17, 2026, DeepSeek launched DeepSeek-V4.1-Flash-1017, centered on 'wish-style programming,' with web, app, and API updated in sync; the company said pricing is unchanged and weights are open. The source says the Search and Fetch APIs each charge 5 yuan per 1,000 calls, while server-side calls within Harness, Claude Code, and Codex begin billing on November 1.
MarkTechPost reported on October 17, 2026 that DeepSeek released the V4.1-Flash-1017 post-training checkpoint under the deepseek-flash model name, claiming improvements across all 11 benchmarks over the September version, with DeepSWE v1.1 at 77.0; it also opened search and fetch APIs to all developers with per-thousand-call billing.
According to an October 16, 2026 MarkTechPost report, StepFun released Step 5 Preview as open weights: 600B total parameters, 27B activated per token, up to 1M context, with FP8 and BF16 checkpoints available on Hugging Face and ModelScope.