Qwen 3.8 27b is still grinding on this!
The threejs demo is still A LOT better...
0.581. That is what the q4_0 KV cache does to long documents, and long documents are the only reason anyone turns it on.
Yesterday I posted the flags that fit a 262K context on a 24GB card: -ctk q4_0 -ctv q4_0 -fa on. Three people had corrected my fp16 table and they were right,
Grok 4.7 is coming… and it’s about to rewrite the AI leaderboard.
I’ve been watching the space closely. Grok 4.6 already pushed hard on pure reasoning and uncensored truth-seeking. But 4.7? Here’s what I’m genuinely expecting — and why it should clear Grok 4.6, Gemini Flash
Qwen 3.8-27B + deepseek V4 pro + V4 flash are free on orcarouter
The free stack All three models live under the -free suffix on Qwen3.8-27B (open-weight, 262K) + DeepSeek V4 Pro + DeepSeek V4 Flash One signup, one key, zero dollars and no card
You have reached the end of the archive
All of qwen38