Model Radar: Six Releases We're Tracking This Week
Our weekly sweep of the model landscape — two open-weights flagships, a multimodal enterprise play, and three router additions worth knowing about.
Every week our discovery pipeline sweeps Hugging Face and OpenRouter for new models worth your attention. Here's this week's shortlist — all six are live on the model radar.
#The flagships
Llama 4 Scout — Meta's open-weights MoE with a 10M-token context window. The release of the week, covered in depth here.
DeepSeek V4 — reasoning-first MoE posting frontier-class math and code numbers at budget inference prices. Our deep dive breaks down the pipeline.
#The enterprise play
Mistral Large 3 — multimodal flagship aimed squarely at enterprise agents, with native document and image understanding and open weights.
#The workhorses
Qwen3.5 72B Instruct — the dense model quietly winning multilingual tool-use evaluations. Available on both Hugging Face and OpenRouter.
GPT-5.2 — OpenAI's latest, now routed through OpenRouter with a 400K window.
Gemini 3 Flash — Google's low-latency multimodal option with 1M context, built for high-throughput agent fleets.
#The pattern
Two things stand out this week: context windows are the new headline spec, and every major release now has an inference-economics story attached. The radar will keep scoring both.
// share this dispatch
// the signal
One email. The week's sharpest AI analysis.
Every Friday: the dispatches that mattered, the models that shipped, and the one chart you need to see. No spam, no filler — unsubscribe anytime.
Join 2,400+ researchers and engineers. Unsubscribe anytime.