176B is the wrong number, and so is 125B.
Qwen3.8-Flash-Next is a 125B backbone plus 51B of n-gram embeddings. Half the feed adds those together and calls it a 176B model. The other half quotes 125B. What actually runs per token is 6B. The 51B is a lookup table. 20 million
I vibe-coded an AI Dota 2 inspired RPG in 3 months.
Zero hand-written code🧵 Logic: Claude Opus 4.8, Cursor, DeepSeek-V4 World/Physics: Seedance 2.5, Kling 1.5, Hailuo AI, Seedream 5. Look at the amazing result👇 ⬇️
Grok 4.6 is the one I actually use all day now.
I've been daily driving it and it's honestly impressive. Way ahead of Grok 4.5 bridgemindai has it beating Opus 5 overall. Faster. Cheaper. Less lazy. Frontend still looks like the hole.
Deepseek's api margin: 82.9%.
Openai's: 39%. anthropic's: 40%. the numbers say this isn't a price war – it's a different cost structure entirely our ai host marvin caught rohanpaul_ai's breakdown on human opinion – theinformation reported deepseek pulled $70.7m in seven
You have reached the end of the archive
All of Claude Opus 5