Qwen releases Qwen3.8-2.4T-A95B, its largest open-weight MoE

A guest post on Terence Tao's blog introduces oai-math-revised, a GitHub repo and browsable site for improving OpenAI's math write-ups, which the authors say can be hard to understand and have incomplete citations. The math community is invited to volunteer as admins and reviewers by email.

A ComfyUI workflow scales photos to 960 px tall on a 2240x960 (21:9) canvas, pre-filling empty sides with mirrored photo edges to stop a grey edge bleeding through the VAE as a faint line. A soft mask lets Qwen paint the sides plus a 224 px band.

Community LoRA aims for phone-camera realism — real skin texture, harsh flash, grain, awkward framing — and merges v1 into a single file with no trigger word. Strength reportedly works from 1.0 upward.

The open-source Linux kernel module unlocks full BF16/FP16 FMA rates on GA102/103/104/106 GPUs without vBIOS flashing; one RTX 3090 test showed bf16 tensor throughput rising from 82 to 155 TFLOPS. Users report up to 80% gains on RTX 30 cards, though prefill/decode gains vary widely.

A Gartner survey found just 13% of CIOs say their enterprise has realized significant or transformational value from AI. The press release is dated October 19, 2026.

On Lenny's Podcast, Anthropic's Emmanuel Ameisen describes the moment he realized Claude was better at SQL than he was, and discusses what happens when AI beats the expert.