I heard claude models are of trillions of parameters !!! 284B, 1M... I wouldn't trust on long running autonomous agents! But for the API costs, this is justified if a better hardware with bigger model is used and will be a claude equivalent for comparable quality at long context retrieval on long running autonomous tasks. At least for me that is important.
I agree that Claude is better (definitely better than the flash version which is relatively small). But...
I actually canceled my Claude Code plan a few months back after trying out some of the "lesser" models on openrouter. They seem to work as just as well (or just as bad) for my coding tasks.
Define "much worse". I use DS v4, GLM, and some Kimi with omp personally, and have Cursor with latest Claude and GPT models at work. I notice zero difference in the work for my workflow between Opus and DS.
Really confused how people make these claims. Are you just basing this off benchmarks or your own personal work? Are you an experienced dev or just doing vibe coding?
My own experience. I'm working on something complex that's not in the datasets these models were trained on. There I see V4 flash breaking down and hallucinating much more often than GPT/Claude. For normal, common tasks, I also don't see much of a difference.
Huge variation in how people prompt and use their models. Vibe coding with ambiguous requirements vs. multiple steps of precise planning are completely different imo
From HF: 284B parameters (13B active), 1M context window.
This is indeed some kind of compressed context and the quality goes down as the context grows. IIRC the V4 paper had some numbers on this
https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash