DeepSeek V4.1 Flash — the best card, 512 GB and a third-party 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 some of the strongest numbers in the index: Terminal-Bench 2.1 at 90.6, DeepSWE v1.1 at 74.2 and a Codeforces rating of 3471, with GPQA Diamond at 90.9, all at maximum reasoning effort. The index scores its capability at 50, the ceiling.
On 17 October DeepSeek published DeepSeek-V4.1-Flash-1017, which its card calls a minor update; the architecture is unchanged. DeepSWE v1.1 rises from 74.2 to 77.0, both on the Mini-SWE-V2 harness. Terminal-Bench 4.0 rises from 31.2 to 42.3, but that comparison crosses harnesses: the 1017 card ran it in Claude Code, while the original card’s 31.2 came from DeepSeek’s own harness. Read the coding gain as real but smaller than the Terminal-Bench jump suggests.
At 4-bit it needs a 512 GB Mac Studio and a third-party runtime. The mainstream runtimes were built for decoder-only models, and none of them handles the encoder-decoder split and its Engram memory tables yet: upstream llama.cpp support is an open pull request (#28696), and stock mlx-lm’s V4.1 module rejects checkpoints that use Engram layers, as the released ones do. oMLX has run it since 11 September, and Rapid-MLX since 19 October (see the updates above). Its 4-bit oQ4e build of the 1017 snapshot is 432 GB, which only a 512 GB Mac Studio holds; an oQ3e build that keeps the Engram tables on SSD is designed for 256 GB machines, and with experts also offloaded to SSD it measured 14.4 tok/s on a 256 GB M3 Ultra (community). A community fork of antirez’s ds4 engine streams a 341 GiB 2-bit GGUF from SSD on a 128 GB M5 Max at 29.6 tok/s (community). Speeds that depend on reading weights from SSD are a different workload, so the index lists that run but never uses it as a Mac speed.
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 — in the index, its Pro sibling and Meta’s Muse Spark Open also take all four natively (Gemma 4 E2B, E4B and 12B take audio and read video as frames). 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. On 27 September Xiaomi published MOPD checkpoints of both V2.6 models, which supersede the RL release: an on-policy distillation stage from several specialist teachers, aimed partly at agents that repeat the same tool call. The scores above are the RL checkpoint’s.
It is a Mac Studio model, and for now not a 128 GB one. The official ggml-org GGUF of the RL release came 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. The RL release’s 2-bit file left a 128 GB Mac no room to work. ggml-org’s GGUFs of the MOPD checkpoint (28 September) add a 105.9 GB Q2_K_S, which a 128 GB Mac can hold, but no Mac run of it has been published, so the index still starts the model at 192 GB.
| Mac | Build | Index speed |
|---|---|---|
| Mac Studio M2 Ultra, 192 GB | MXFP4 / MLX | estimated~54 tok/s |
| Mac Studio M3 Ultra, 256 GB | MXFP4 / MLX | community59 tok/s |
| Mac Studio M5 Ultra, 256 GB | MXFP4 / MLX | community62 tok/s |
When this was written the table used the index’s MoE factor, which this week’s first measurements of a sparse model on the Ultras suggested was conservative — see how the estimates held up. It is now refilled from the index on every build: measured where a run exists, otherwise scaled from the measured M3 Ultra runs, with the mark on each figure saying which. 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 (1017 snapshot: 17 Oct) | 21 Sep 2026 (MOPD: 27 Sep) | 21 Sep 2026 (MOPD: 27 Sep) |
| 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 | 512 GB Mac Studio at 4-bit (oMLX or Rapid-MLX) | 192 GB Mac Studio | None — server-class |
| Index speed, M5 Ultra | estimated~29 tok/s | community62 tok/s | — |
Every benchmark figure here comes from the vendor’s own card and harness; independent standardized runs usually come in lower.
What to watch
- Upstream llama.cpp support (PR #28696). oMLX and Rapid-MLX already run V4.1 Flash on a Mac Studio; upstream support would make the GGUF builds usable too. #28696 was still open on 24 October, and a competing draft, #30817, targets the 1017 checkpoint. (Updated 24 October 2026.)
- A measured 128 GB MiMo run. ggml-org’s 105.9 GB Q2_K_S of the MOPD checkpoint (28 September) fits a MacBook Pro or Mac Studio M5 Max with 128 GB; a published run of it would bring MiMo-V2.6-Flash down a tier. None had been published by 24 October. (Updated 24 October 2026.)
- Measured runs on the M5 Ultra. Closed for MiMo-V2.6-Flash: its 61.7 tok/s community run replaced the scaled estimate and moved the index’s #1 (see the update above). DeepSeek V4.1 Flash still has no 4-bit M5 Ultra run, so its M5 Ultra figure is still scaled from the M3 Ultra runs. (Updated 24 October 2026.)
Frequently Asked Questions
Can I run DeepSeek V4.1 Flash on a Mac?
Yes, on a 512 GB Mac Studio at 4-bit, through oMLX, an MLX-based runtime that has supported it since 11 September 2026, or Rapid-MLX 0.16.5 and later. An mlx-community oQ4e build of the 17 October '1017' snapshot is 432 GB; oMLX 0.7.2 runs it at 38.7 tok/s on a 512 GB M3 Ultra (MTP off, 16K-token prompt; community), against 20.23 tok/s for an earlier build at a 4K prompt. A 256 GB Mac runs it only below 3 bits or with experts offloaded to SSD. As of 24 October 2026 upstream llama.cpp support is still an open pull request (#28696, with a competing draft, #30817), and stock mlx-lm does not load the released weights, which use Engram memory layers.
What Mac do I need for MiMo-V2.6-Flash?
A 192 GB Mac Studio or larger, by the LLMCheck index. The MXFP4 GGUF is 167 GB and the mlx-community MLX build loads in about 170 GB. ggml-org's 105.9 GB Q2_K_S of the MOPD checkpoint fits a 128 GB Mac, but no Mac run of it has been published. Community runs measured 55 to 59 tok/s on an M3 Ultra and 61.7 tok/s on a 256 GB M5 Ultra (4-bit, MTP off); the M5 Ultra run replaced the index's earlier estimate of about 79.
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; the MOPD checkpoints that superseded them on 27 September are MIT too. 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 (77.0 for the 17 October '1017' snapshot). MiMo-V2.6-Flash accepts audio and video as well as images and needs only a 192 GB Mac Studio, where V4.1 Flash needs a 512 GB one at 4-bit and runs only through third-party MLX runtimes. As of 24 October 2026 the LLMCheck index scores MiMo-V2.6-Flash 78 (#7) and V4.1 Flash 70; the index's #1 is Qwen 3.6-35B-A3B (Score 81), which needs only 24 GB.
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.