Friday, August 28, 2026

llama.cpp can run Qwen3.8-Flash-Next. Newer Macs get faster Metal.

llama.cpp now runs Qwen3.8-Flash-Next on its own, with downloadable GGUFs from Unsloth. This morning it also shipped faster Metal attention for M4 Pro, M3 Max, M5, and M5 Pro. An M3 Ultra does not automatically get the Max tables. Build numbers, file sizes, and the mlx-lm leak fix are in Local and MLX. Elsewhere: Bedrock AgentCore in two more regions, Claude Console keys, MiniMax-H3 on H200s (vendor), and GLM-5.3-Flash MIT weights. From the field: OpenExecutive, Ante, and Arcee NAC. NVIDIA/HF still unconfirmed. No frontier flagship.

Window ~2026-08-01 to 2026-08-28. Thursday already covered NVIDIA/HF reports, Q2 $96.2B, AWS 2M GPUs, Ollama v0.33.1 MLX Flash Next, llama.cpp b10646 Metal leak, mlx-vlm 0.6.17, SGLang day-0 Qwen Flash Next. Primary sources only. Vendor evals flagged.