Signal
12 articles covering "open-weight models"
Anthropic reports Z.ai's open-weight GLM-5.3 developed end-to-end exploits in 50 of 410 ExploitBench attempts versus 56 for Claude Mythos Preview; NIST's CAISI calls it the most cyber-capable open-weight model.
Qwen-Image-2.1 unifies text-to-image generation and editing with native RGBA transparency, tops open-weight rivals on Qwen's own benchmark, but trails closed leaders.
Tencent's newly open-sourced Hy4 preview model jumped to 8th place on the Code Arena WebDev leaderboard, up from 34th for its predecessor, while outscoring Qwen3.8-Max and DeepSeek-V4 Pro on the DeepSWE benchmark.
Z.ai released GLM-5.3-Flash under an MIT license, confirming it was the anonymous 'ox-alpha' model that had been topping OpenRouter and OpenCode leaderboards.
Alibaba's Qwen team shipped Qwen3.8-27B as Apache 2.0 open weights, the on-premise checkpoint promised alongside Qwen3.8-Max, with coding benchmarks that top Claude Opus 4.6 Max on two of four tests.
GitHub made Moonshot AI's Kimi K3 generally available in Copilot on August 6, briefly pausing the rollout to mitigate a GitHub Actions incident before resuming it the same day.
Nvidia's open-weight Nemotron 3.5 Lightning targets the repetitive execution layer of AI agents, launching alongside a NeMo Switchyard library that routes tasks between it and frontier models.
Meta released Muse Glimmer, a 30B-parameter open-weight model distilled from Muse Spark for local agent workflows, alongside a Zuckerberg essay on open-source AI.
Alibaba made its new flagship Qwen3.8-Max, a 2.4 trillion-parameter multimodal model, broadly accessible ahead of an open-weights release, sending its Hong Kong-listed shares up 7 percent.
Moonshot AI's Kimi K3, priced far below Western rivals, triggered a Friday selloff in Chinese AI and global chip stocks before demand forced a subscription pause.
Thinking Machines Lab shipped its first model, Inkling, an open-weight 975-billion-parameter system that tops US open-weight benchmarks but still trails China's leading models.
Shanghai-based MiniMax launched M3 on June 1, pairing a 1-million-token context with a new sparse-attention design and company benchmarks that top GPT-5.5, with weights promised within 10 days.