Local and open models are becoming a real daily option

antirez's ds4 got a warm reception. Commenters are already forking the idea: one wrote an inference engine for Intel Xe-LP laptops running quantized Gemma-4, another got 1M context on a 128GB M5 Max using Qwen. In the Ask HN about replacing Claude or GPT locally, one user said Qwen 3.6 27b dense feels about equal to Haiku 4.5 and maybe Sonnet on some tasks. Another runs DeepSeek V4 Flash on two RTX Pro 6000s at 160 tok/s. Germany's Kolibri sovereign model and the FLUX 3 open-weight release add to the same pull.

The pattern: open weights are no longer just a hobby. They are good enough for the middle of the work, such as drafting, boilerplate and overnight batch jobs. The key bit is that people don't treat this as all or nothing. They route cheap tasks to local or open models and keep the frontier model for the hard parts.

The counterpoint is real. Tokens per second on a laptop still lag the cloud, aggressive quantization hurts quality (one commenter said the quantized DeepSeek checkpoint isn't very good), and there is almost no enterprise tooling for picking and operating models. Kolibri got a lukewarm reception for overthinking and weak tool calling.