DeepSeek V4.1 Flash — the best card, no Mac runtime
DeepSeek published DeepSeek-V4.1-Flash on 10 September under MIT. It is not a refresh of V4 Flash. The card describes a causal encoder-decoder: a 40-layer transformer split into 20 encoder and 20 decoder layers, with a mixture-of-experts feed-forward of 384 routed experts — six fire per token, alongside one shared expert. That comes to 552B parameters in the backbone, 8B active per token while reading the prompt and 16B while generating, with image input and a 1M-token context.
Its card carries the strongest numbers in the index: Terminal-Bench 2.1 at 90.6, DeepSWE v1.1 at 74.2, GPQA Diamond at 90.9 and a Codeforces rating of 3471, all at maximum reasoning effort. The index scores its capability at 50, the ceiling.
It does not run on a Mac yet. llama.cpp, MLX and Ollama all expect decoder-only models, and none implements the encoder-decoder split as of 23 September. GGUF files have appeared under individual accounts, but upstream llama.cpp cannot load the architecture, so treat them as experiments. When support lands, a 4-bit build would need around 315 GB — a 512 GB Mac Studio — and 16B active parameters per generated token would put it in the same speed class as DeepSeek V4 Flash. Until then the index lists it with no speed figure.
MiMo-V2.6-Flash — omnimodal, from 192 GB
Xiaomi’s MiMo team published MiMo-V2.6-Flash on 21 September, ungated and MIT-licensed. From the card: 309B parameters with 15B active, a hybrid of sliding-window and global attention, a five-layer multi-token-prediction head, a 1M-token context, and native text, image, video and audio input — it and its Pro sibling are the only models in the index that take all four. The card reports Terminal-Bench 2.1 at 87.6, DeepSWE v1.1 at 67.9 and Toolathlon-Verified at 73.6, which puts it level with GLM-5.3; the index scores it 49.
It is a Mac Studio model, and not the 128 GB kind. The official ggml-org GGUF comes in two sizes, Q2_K at 126 GB and MXFP4 at 167 GB; the mlx-community build is 156 GB on disk and about 170 GB once loaded. Even the 2-bit file leaves a 128 GB Mac no room to work.
| Mac | Build | Estimated speed |
|---|---|---|
| Mac Studio M4 Ultra, 192 GB | MXFP4 / MLX | ~39 tok/s est. |
| Mac Studio M3 Ultra, 256 GB | MXFP4 / MLX | ~30 tok/s est. |
| Mac Studio M5 Ultra, 256 GB | MXFP4 / MLX | ~43 tok/s est. |
These use the index’s MoE factor, which this week’s first measurements of a sparse model on the Ultras suggest is conservative — see how the estimates held up. The MTP head matters too: runtimes that support multi-token prediction can draft with it, and the GGUF release ships it as a separate file.
MiMo-V2.6-Pro — the 1T flagship
MiMo-V2.6-Pro shipped the same day: 1.02T parameters with 42B active, the same four input types and 1M context, also MIT. Its card reports Terminal-Bench 2.1 at 89.9 and DeepSWE v1.1 at 71.9. At about 580 GB in 4-bit it is a server model; the index lists it for completeness, without a Mac speed.
Side by side
| DeepSeek V4.1 Flash | MiMo-V2.6-Flash | MiMo-V2.6-Pro | |
|---|---|---|---|
| Weights published | 10 Sep 2026 | 21 Sep 2026 | 21 Sep 2026 |
| Parameters / active | 552B / 16B decoding | 309B / 15B | 1.02T / 42B |
| Input | Text, image | Text, image, video, audio | Text, image, video, audio |
| License | MIT | MIT | MIT |
| Terminal-Bench 2.1 (card) | 90.6 | 87.6 | 89.9 |
| Smallest Mac today | None — no runtime | 192 GB Mac Studio | None — server-class |
Every benchmark figure here comes from the vendor’s own card and harness; independent standardized runs usually come in lower.
What to watch
- Encoder-decoder support in llama.cpp or MLX. The first runtime that loads V4.1 Flash turns the strongest open model into a 512 GB Mac Studio model.
- A 128 GB MiMo build. A 1-bit or dynamic 2-bit quant under about 110 GB would bring MiMo-V2.6-Flash to the MacBook Pro M5 Max. None exists as of 23 September.
- Measured runs on the Ultras. Both Flash models carry estimates only; the Mac Studios shipped on 22 September.
Frequently Asked Questions
Can I run DeepSeek V4.1 Flash on a Mac?
Not as of 23 September 2026. V4.1 Flash uses a new causal encoder-decoder architecture that llama.cpp, MLX and Ollama do not support yet, and GGUF uploads from individual accounts cannot be loaded by upstream llama.cpp. When support arrives, a 4-bit build would need roughly 315 GB, which only a 512 GB Mac Studio can hold.
What Mac do I need for MiMo-V2.6-Flash?
A 192 GB Mac Studio or larger. The smallest official GGUF is Q2_K at 126 GB, too large to leave working room on a 128 GB Mac, and the MXFP4 GGUF is 167 GB; the mlx-community MLX build loads in about 170 GB. The LLMCheck index estimates about 39 tok/s on an M4 Ultra and 43 on an M5 Ultra.
Is MiMo-V2.6 open source?
Both MiMo-V2.6-Flash and MiMo-V2.6-Pro were published on 21 September 2026 with open weights under the MIT license on Hugging Face, ungated, with technical reports. MIT permits commercial use, modification and redistribution with attribution.
Which is better, DeepSeek V4.1 Flash or MiMo-V2.6-Flash?
On their own cards V4.1 Flash scores higher: Terminal-Bench 2.1 90.6 against 87.6, and DeepSWE v1.1 74.2 against 67.9. MiMo-V2.6-Flash accepts audio and video as well as images and runs on a 192 GB Mac Studio today, while V4.1 Flash has no Mac runtime yet. For a Mac owner in September 2026, MiMo is the one you can use.
Does MiMo-V2.6-Flash support MTP?
Yes. It ships a five-layer multi-token-prediction head, and the official GGUF release includes it as a separate file. Runtimes that support MTP speculative decoding can use it to generate faster; the LLMCheck estimates assume plain decode and do not include that gain.