If 5090 has 32GB, and let's say somehow a 1-bit quantization is possible and you don't need more VRAM for anything else (forget KV cache etc), it would be able to fit a 256B 1-bit model. Just to picture it in extremes how unlikely this is.

And the active parameters come from the experts. For each token the model picks some experts to run the pass (usually 2 to 4, I haven't read V4's papers). It's not always the same experts.

OTOH, being DeepSeek, I foresee a bunch of V4 distilled FP8 models fitting in a 5090 with tiny batches and with performance close from 75 to 85% of V4. And this might be good enough for many everyday tasks.

Today is a good day for open models. Thank god for DeepSeek.

Theoretically with streaming, any model that fit the disk can run on consumer hardware, just terribly slow.

If that is really so, it would be now be good enough to replace claude for us; we use sonnet only; with our setup, use cases and tooling it works as well as opus 4.6, 4.7 so far. We won't replace sonnet as long as they have subscriptions but it is good to have alternatives for when they force pay per use eventually.

Models like Deepseek is the only reason we are able to categorize and measure quality of thousands of MCP servers (https://glama.ai/blog/2026-04-03-tool-definition-quality-sco...). That's billions of tokens – an expense that would be otherwise very hard to swallow.

I estimated that even with heavy usage it would cost your around 30-70$ depending on caching at around 40M tokens. That would give you around double the usage compared to gpt-5.5 on the 200$ sub

This is refreshing right after GPT-5.5's $30

I believe the R stood for reasoning, just like OpenAI had their own dedicated o1/o3 family, but now every model just has it built-in.