Yes — with 73.0 GB to spare
DeepSeek V4.1 Flash 552B-A16B at Q4_K_M fits your M3 Ultra · 512 GB entirely in unified memory at 8K context, at an estimated 23 tokens per second. There is room for its full 1024K window.
Twice V4 Flash's backbone (552B, 16B active per generated token) plus 196B of Engram lookup tables (189 GiB at FP8) that DwarfStar streams from SSD, so they are not counted here. A 512 GB Mac model, and as of October 2026 no mainline llama.cpp or Ollama build: DwarfStar on a Mac, vLLM across four GPUs. DeepSeek puts the cache at 890 bytes/token; modelled conservatively as 4 latent layers plus a 128-token window.
What hardware do I need for DeepSeek V4.1 Flash 552B-A16B? →
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 546.2 GB | 546.9 GB | — | ~13 | −0.1% ppl | 162.9 GB over |
| Q6_K | 421.6 GB | 422.2 GB | — | ~17 | −0.4% ppl | 38.2 GB over |
| Q5_K_M | 364.4 GB | 365.0 GB | 1024K | 20 | −0.8% ppl | Long context |
| Q4_K_M | 310.4 GB | 311.0 GB | 1024K | 23 | −1.9% ppl | Recommended |
| Q3_K_M | 251.3 GB | 251.9 GB | 1024K | 29 | −5.4% ppl | Long context |
| Q2_K | 215.3 GB | 215.9 GB | 1024K | 33 | −15% ppl | Long context |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it. This model interleaves sliding-window layers (128 tokens, 4 of 40 layers global), which is why its cache barely grows with context. This model uses multi-head latent attention, so its cache is a compressed latent rather than full K and V.
How to run it
# in a DwarfStar checkout on the ds4.1flash branch $ ./download_model.sh ds41f-q4 $ ./ds4 -m gguf/DeepSeek-V4.1-Flash-Q4.gguf --ctx 8192
On a Mac with less than 512 GB, download ds41f-q2 instead and add --ssd-streaming.
antirez's Metal inference engine for DeepSeek V4, with V4.1 support on its ds4.1flash branch. More on DwarfStar.