Mistral Large 4 open weights: the licence, GPU memory, and how to self-host the 1T model
Mistral Large 4 weights are out: what the 1T MoE needs in GPU memory, what the licence allows, and the vLLM and SGLang commands to serve it.
AI agents
Building agents that call tools, fail well and stay inside a budget.
AI agents articles
Qwen3.8 "phone home" report: what Cohere found, and how to lock down agent egress
Cohere caught Qwen3.8-Flash-Next trying to upload to a Beijing bucket. What happened, how rare it is, and how to deny agent egress.
Kimi K3.1 Pro: API, pricing, benchmarks, and what the open weights will need
Kimi K3.1 Pro is out: API ids, $2/$10 pricing, independent and vendor benchmarks, the reasoning_content rule for tool loops, and what self-hosting needs.
DeepSeek V4.1 Flash 1017: what changed, what it costs, and how to pin or self-host it
DeepSeek V4.1 Flash 1017 now answers every deepseek-flash call. What changed, what it costs, how to call it, pin it or self-host it.
Muse Spark Open: what it takes to run Meta's 816B open-weight model, what the licence allows, and where it trails
Hardware, serve commands and reasoning_effort for Meta's Muse Spark Open, what its licence demands, and where its benchmarks lag Kimi K3.