Open weights keep closing the gap, with licensing strings
Mistral released the weights of Mistral Large 4, a 1T-parameter model with 52B active (50289418). Commenters immediately pointed out it's weights-available, not open: it uses a modified MIT license that requires a commercial license if you serve it as a business. One user complained it's weirdly hard to even give Mistral money. Meanwhile an Ask HN (48542100) asks if anyone has replaced Claude or GPT for daily coding. Answers: Qwen 3.6 27B dense feels like Haiku 4.5, maybe Sonnet on some tasks. One person runs DeepSeek V4 Flash on 2x RTX Pro 6000 Blackwell at 160 tok/s. Others say Gemma 4 on an M4 is too slow.
The key bit: local is good enough for a slice of work, but speed and tooling still lose to the cloud. Artificial Analysis v5 (50273340) adds another wrinkle. The writeup argues Anthropic and OpenAI stopped caring about these public indexes months ago, and commenters want benchmarks run at the reasoning settings people actually use, not 'max'.
The nuance: the benchmarks measure what open labs optimize for, not what closed labs ship, so leaderboard parity may overstate real parity.
So what?
Open models are a real fallback and a negotiating chip against API vendors. Read the license before building a business on 'open' weights, because 'serving it commercially' is exactly what you'd be doing. Test models on your own tasks and ignore the leaderboards.