Filed source · tool
GLM-5.3
z.ai
Z.ai's next model reuses the GLM-5.2 base and puts every gain into post-training — weights held back roughly two weeks this time, after training produced cybersecurity capabilities the team didn't plan for.
Visit original sourceFiled under
Related entries
Seed 2.1 Turbo
seed.bytedance.comtool
ByteDance's low-latency tier of Seed 2.1, on Volcano Engine — pitched as beating Claude Opus 4.6 at up to 80% lower cost for high-frequency agent workloads.
DeepSeek-V4-Flash-0731
huggingface.cotool
DeepSeek retrains its flash tier of V4 for coding, agents and tool use — 284B total parameters, 13B active, and a 1M-token context window.
Qwen3.7 Flash
help.aliyun.comtool
Alibaba's cheap, fast vision-language tier of Qwen3.7 — $0.03/M input tokens, a 991K context window, and no accompanying technical report.
Kimi K3
huggingface.cotool
Moonshot open-sources a 2.8-trillion-parameter, 1M-context model — billed as the largest open-weight model released so far, with only 104B parameters active per token.