Open-weight models keep closing the gap, and people want them local

DeepSeek v4.1 Flash 1017 landed in two separate threads (50129121, 50129045). One commenter asked whether the benchmarks could possibly be real, another noted it sits above GLM 5.3, and a third pointed out that a "minor update" showed substantial gains. Ultramarine 1 Flash (50106747) is a professional-knowledge-work agent built by running large-scale reinforcement learning on top of the open GLM-5.3 Flash base. Apple published a post on running Muse Spark Open, a nearly trillion parameter model that never mentions Meta in the text, on a Mac Studio with MLX (50118320).

The hardware side is just as active. Llama.cpp now runs on a jailbroken PS5 with 11.4G of shared memory (50122431). Llmspeed (50127979) lets you feel local model speed before buying hardware, and commenters immediately argued its numbers were off, since an M2 Ultra reportedly runs faster than the listed M3 Ultra. An older Ask HN on replacing Claude or GPT with local models (48542100) shows the same mix: Qwen 3.6 27B described as roughly Haiku 4.5 level, DeepSeek V4 Flash on two RTX Pro 6000s at 160 tok/s, and people still giving up on M4 laptops because tokens per second are too low.

The key bit: the open ecosystem now ships frontier-ish models faster than most people can host them. The bottleneck has moved from model quality to memory, tooling and honest measurement. Commenters called out Apple's post for benchmarking Meta's model at 4-bit against Qwen at BF16 only, a reminder that vendor charts are marketing.