本地AI真的起来了，27B的Qwen能打赢405B的Llama，GLM 5.2还能压到250GB，普通显卡就能跑。
Qwen 27B在RTX 3090上超过了21个月前发布的Llama 405B。GLM 5.2通过保护超权重，从1.5TB压缩到250GB。Cognition用前沿模型规划、廉价模型执行，将Fable级智能成本降低40%。OpenRouter的自动路由在两年后因openclaw的十分钟心跳找到用途。EXO Labs与NVIDIA合作三周，让DGX Spark提速10倍。
Local is no longer the compromise: - A dense 27B Qwen now beats the 405B Llama from 21 months earli...
Local is no longer the compromise: - A dense 27B Qwen now beats the 405B Llama from 21 months earlier, on the same RTX 3090 that used to host Llama 2 - Quantize one wrong number, just one, and a model gets 20% dumber. Respect the super weights and GLM 5.2 shrinks from 1.5TB to 250GB - Cognition cut the cost of Fable level intelligence by 40% by letting the frontier model plan while cheaper models implement - OpenRouter's auto router sat almost unused for two years until openclaw's ten minute heartbeats found it a purpose - Three weeks of NVIDIA swarming with EXO Labs got a 10x speedup on the DGX Spark without solving any new computer science Watch the Local AI playlist: youtube.com/playlist?list=… 💬 0 🔄 0 ❤️ 1 👀 413 📊 1 ⚡