DeepSeek brings search, file, and URL tools to its Flash API
DeepSeek added web search, URL reading and file input to its V4.1-Flash API. Here is what retrieval costs and how to test it.
DeepSeek, Kimi, Qwen, and Chinese models reshaping the race.
DeepSeek added web search, URL reading and file input to its V4.1-Flash API. Here is what retrieval costs and how to test it.
Alibaba released the stable Qwen4-Max API. Compare model IDs, access, and output limits before the preview ends on November 25.
Baidu reports 14.9% faster decoding in one optimized setup and 6.236 times baseline throughput in a separate low-latency P800 test.
Axion 1.0 shares training and data-construction code. Read the paper’s evaluation limits and distinguish the linked community checkpoint from official weights.
Baidu releases ERNIE 5.5 Flash Base weights. The checkpoint needs adaptation, and its custom license includes territory and commercial-use conditions.
DeepSeek’s new repository shares a proactive coding-agent prototype with authorization rules. Its authors explicitly warn that it is not production-ready software.
Meituan offers LongCat 2.5 through its API and as MIT-licensed weights. The main and multimodal weight files total about 3.56 TB.
Tencent releases Hy-SimulMT for live Chinese-English translation. Review model variants, separate speech output and the documented deployment limits.
Tencent publishes Hy-Nomas in BF16 Diffusers format. Check its pipeline, retained history and camera control; 24 FPS describes playback rather than generation speed.
Together AI hosts Kimi K3.1 Pro with a 1M-token context. Compare token prices, post-training limits and the separate public-weights deployment route.
DeepSeek’s DeepInfer release supports offline inference on specified NVIDIA and Ascend hardware. Review the missing service components before planning deployment.
MiniMax M3.1 Flash is available through its API, M Plan and MiniMax Code. Check published token rates, completed-task costs and the status of model…