Llm

A concise 1-page 5W1H breakdown of Mistral Large 4 ("le Chonk"): who built it and who it impacts, the 5 core highlights (1T/52B MoE architecture, hybrid reasoning & agents, #1 OSS defensive cybersecurity, multimodal STEM/CAD grounding, and 33B tokens/day RL), benchmarks, and Hacker News takeaways.

A concise 1-page 5W1H breakdown of DeepSeek-V4.1-Flash-1017: 763B MIT-licensed multimodal MoE, 77.0% on DeepSWE v1.1, 42.3% on Terminal-Bench 4.0 (+11.1 pts), the Rust deepseek-recipe SDK, and why devs pair it with Opus 5.5 or Sol 6.1.

A concise 1-page 5W1H engineering brief on Anthropic Claude Haiku 5.5: 90% lower per-token rates (≤100K), 1M context window, 72.4% on OSWorld 2.1, head-to-head benchmarks against Haiku 4.5, GPT-6 Luna, and Sonnet 5.5, plus breaking API migration rules.

A concise 1-page 5W1H breakdown of Meta AI’s Muse Spark Open (816B total / 42B active MoE): 1M context, Helion GPU kernels, 4x M5 Ultra Mac Studio MLX benchmarks, the $20M MaaS license cap, and why Hacker News compared it against GLM, Qwen, and DeepSeek.

A concise 1-page 5W1H breakdown of Google DeepMind’s Gemini 4 Argon: 1M input/output context, the December 1 price doubling ($2/$10 to $4/$20), 5.8x effective coding token cost vs. Sol, and why Hacker News praised its prose while roasting its rollout.