DeepSeek V4.1 Flash runs at 20 tok/s (20.23 as recorded) on a 512 GB M3 Ultra, a community run reported by github.com (MLX, oQ4e, 4K context). It needs about 434 GB of unified memory. It is DeepSeek's 552B mixture-of-experts model (16B active), released under the MIT licence.
Only a third-party runtime runs it: oMLX's own deepseek_v41 support, on a 512 GB Mac. Upstream llama.cpp support is still an open PR (#28696).
This is a community run: a real run, reported by its source and linked below. The index records it as given and never adjusts it.
| Metric | Value |
|---|---|
| Tokens per second | community20 tok/s |
| Provenance | community run · github.com |
| Quantization | oQ4e |
| Engine | MLX |
| Context | 4K |
| Figure dated | 2026-09-11 |
| Minimum RAM | 434 GB |
| Parameters | 552B MoE (16B active) |
| Chip bandwidth | 819 GB/s |
Each row as the index holds it, with its source. “Mac RAM” is the memory of the Mac a run used, not the model’s requirement (434 GB). The headline figure follows the reference rule: the lower-median measured plain-decode run at 5 bits or fewer (a run whose quant was not stated counts only when no stated-quant run qualifies), else the estimate; a run marked “not used to anchor estimates” is listed here but never becomes the headline.
| Speed | Quant | Engine | Mac RAM | Context | Source | Note |
|---|---|---|---|---|---|---|
| community20.23 tok/s | oQ4e | MLX | 512 GB | 4K | github.com | Base V4.1 checkpoint (the 1017 snapshot has the same architecture). oMLX 0.7.0.dev2, -mtp checkpoint run with MTP off, Engram tables in RAM. 19.67 tok/s at 64K. |
| community6.53 tok/s | 4/8-bit mixed | MLX | 512 GB | 8K | huggingface.co | PipeNetwork custom V4.1 runtime on mlx 0.32.2 / mlx-lm 0.31.3 (immature loader); 427.6 GB.; not used to anchor estimates |
Ollama — the index has not verified an Ollama tag for DeepSeek V4.1 Flash, so it doesn’t print one (a guessed tag fails with “pull model manifest: file does not exist”). Search the Ollama library →
MLX — pick a build from Hugging Face’s MLX listings, then run it with its repo id:
Downloads: Hugging Face repo → · model card →
The 20 tok/s figure above came from MLX at oQ4e; other runtimes and quants run at different speeds.
Prefer a GUI? LM Studio downloads GGUF and MLX builds in one click. Step by step: install Ollama on a Mac.
| Chip | Speed | Quant | Engine | Bandwidth |
|---|---|---|---|---|
| M5 Ultra | estimated29 tok/s | oQ4e | MLX | 1228 GB/s |
| M5 Max | estimated12 tok/s | 6.6 bpw | — | 614 GB/s |
DeepSeek's MIT-licensed 552B causal encoder-decoder MoE (16B active per decoded token) with image input and 1M context; the 17 Oct 2026 '1017' snapshot lifts DeepSWE v1.1 from 74.2 to 77.0 (Mini-SWE-V2 harness). It runs on a 512 GB Mac Studio through oMLX: the oQ4e MLX build is 432 GB, and a community run of the base checkpoint measured ~20 tok/s on an M3 Ultra with MTP off. Upstream llama.cpp support is still an open PR (#28696), and Ollama is cloud-only.
MIT licence · 552B MoE · 1000K context · LLMCheck capability 50/50 · see its leaderboard rank →
DeepSeek V4.1 Flash needs about 434 GB of unified memory. These current Apple Silicon Macs have the headroom to run it comfortably:
As an Amazon Associate, LLMCheck earns from qualifying purchases. These affiliate links cost you nothing extra and help keep the LLMCheck index free.
See how DeepSeek V4.1 Flash stacks up against other models on your specific Mac hardware.
Open Compare Tool Full Leaderboard