StepFun opened API access to Step 5 Preview on Sept 20: a proprietary 600B sparse MoE (27B active) with a 1M-token context, vision input, and Artificial Analysis pricing of $1/$2.70 per million tokens (Intelligence Index 44). Company-reported coding scores still trail GPT-6 Astra and Claude Opus 5 on several benches; open weights are planned for Oct 15.
Alibaba previewed Qwen Image 2.1 on Sept 20: multi reference conditioning that turns a three view character sheet into a six panel storyboard with clothing and appearance held across scenes. Diffusers docs describe a unified generation/editing pipeline; open weights remain listed as upcoming on ModelScope…
Scalably’s Sept 19 field report: two RTX 5090s ran Unsloth’s Qwen3.8 27B NVFP4 via vLLM 0.27.0 for 14 days of real agent traffic—82.6% prefix cache hits, 28k requests, zero engine errors. A faster MoE lost the job on tool recovery (1/20 vs 12/20), so they kept the dense model…
Open Interpreter 0.0.45 (Sept 20) adds DeepSeek Flash and DeepSeek V4 Pro to the model picker; Flash supports image input and a 1M token context. Also fixes provider picker loading and a UTF 8 boundary issue…
Thomas Ahle used Claude Fable to audit SMHasher fast hashes and found seed independent or weak key collisions in xxHash, komihash, CityHash, MurmurHash3 and others—often far below claimed guarantees. Findings were disclosed upstream; the post argues AI has changed the threat model for heuristic hashes…
Accomplish AI researchers disclosed Heapjack and Overpatch: Codex sandbox escapes that could run host commands, including from read only mode via a shared Node heap token. OpenAI patched both within eight days; update Desktop to 26.818.21641+ and CLI to 0.149.0+…
Spotted something we missed? Start a thread.