Qwen 3.6-35B-A3B Is #1 Again: MiMo-V2.6-Flash Measured on an M5 Ultra

The new top of the leaderboard

The LLMCheck Score adds four parts: capability (0–50), speed (0–25: the reference figure in tok/s × 0.25), accessibility (0–15, more for less memory) and licence openness (0–10). With MiMo-V2.6-Flash’s M5 Ultra speed now measured, the top ten on 24 October reads:

Qwen 3.6-35B-A3B scores 81 = 38 + 25 + 10 + 8. Its M5 Max reference figure, sourced118 tok/s in Liquid AI’s published llama.cpp Q4_K_M baseline, is past the point where speed points cap, and its 24 GB minimum earns 10 accessibility points. Recursive Flint, new this week, is #2 at 80 on an estimated M5 Max figure; the Flint post explains how that estimate is built. MiMo-V2.6-Flash is #7, tied at 78 with Gemma 4 26B-A4B and MiniCPM-V 4.7 and placed between them by name. The live order is always on the leaderboard.

MiMo-V2.6-Flash on an M5 Ultra: the runs

Xiaomi’s MiMo-V2.6-Flash is an MIT-licensed 309B mixture-of-experts with 15B active parameters. Its 4-bit MLX builds take about 155 GB in use, so it needs a Mac Studio with 192 GB or more. These are all the plain-decode runs the index holds:

MiMo-V2.6-Flash: every measured plain-decode run in the LLMCheck index (data version 2026-10-24). All are 4-bit MLX builds of the RL checkpoint with multi-token prediction off.

The figure the index had been scoring was its own estimate, about 79 tok/s at MXFP4, scaled from the two older M3 Ultra runs (59.4 and 59). The eligible M5 Ultra run measured 61.7, 22% below that estimate. It is the second time this month that a measured run has taken #1 away from an estimate, after KAT-Coder-V2.5 on 18 October, and the costlier of the two: four Score points against three. The speed-estimate check now carries it.

All five runs used Xiaomi’s RL checkpoint. The catalog points to the newer MOPD checkpoint (27 September), which has the same architecture, so the index treats the runs as valid for it. Two more M5 Ultra figures are not plain decoding and stay in prose: 101.8 tok/s at a 4K context with oMLX’s Lightning MTP on (community, RL checkpoint, 80-core GPU, source), and 59.8 tok/s at 1K for a MOPD oQ4e build with MTP on (community, on another machine labelled 64-core, source). Multi-token prediction drafts several tokens per step, which is a different workload from the plain decode every leaderboard figure measures.

Why the rule moved it from 82 to 78

One function in the index’s build, the reference-row rule, decides the figure each model is scored on (methodology):

1. The reference chip is the M5 Max, or the Mac Studio M5 Ultra for a model that needs more than 128 GB. MiMo-V2.6-Flash needs 170 GB, so its reference chip is the M5 Ultra.
2. A measurement beats an estimate. Among measured plain-decode runs on that chip at 5 bits or fewer, the index takes the lower median, never the fastest run.
3. Some runs are shown but never count: runtimes patched with unmerged code, runs whose own record is doubtful, SSD-streamed runs and cut-down (binned) chips.

The 76.4 tok/s run falls under the third point. Its page reports a 64-core GPU with 256 GB, a combination Apple does not offer. Memory bandwidth is 1.2 TB/s on either GPU, so the core count alone would not explain a faster figure; the trouble is that the page describes a machine that does not exist, and the index cannot tell which detail is wrong. The row is displayed but does not count. That leaves one eligible run, 61.7 tok/s, displayed as 62.

MiMo-V2.6-Flash’s other three parts did not change: capability 49, accessibility 3 (a 170 GB minimum) and licence 10 (MIT).

MiMo-V2.6-Flash still has the highest capability figure in the top ten, and it is still the index’s pick for every Mac with 192 to 512 GB: community62 tok/s community-measured on the Mac Studio M5 Ultra.

Why the M5 Ultra is barely faster here

On paper an M5 Ultra should run MiMo-V2.6-Flash about 1.5× as fast as an M3 Ultra: 1,228 against 819 GB/s of memory bandwidth, and the old estimate assumed most of that gain. The eligible M5 Ultra run is only 1.04× to 1.12× the three M3 Ultra runs (55.1, 59 and 59.4 tok/s, all community-measured).

Other models do get most of the gain. One tester ran Qwen 3.6-35B-A3B on both Mac Studios in mlx_lm.benchmark: 153.4 tok/s on the M5 Ultra and 106.4 on the M3 Ultra, 1.44× (community-measured). Apple’s ML research team published Qwen3.8-27B at BF16 at 19.69 and 12.46 tok/s, 1.58× (published). So the shortfall looks specific to MiMo on today’s runtimes, not to the chip. The index cannot yet separate two explanations: per-token work in MiMo’s MLX path that does not scale with bandwidth, and early oMLX builds (a v0.7.0 release candidate on the M5 Ultra) on a chip that shipped in September. The unexplained 76.4 run, 1.29× the best M3 Ultra figure, hints that a better-tuned setup exists.

The practical reading: on current software, a 256 GB M3 Ultra Mac Studio runs this model at 89–96% of the M5 Ultra’s measured speed. The M5 Ultra vs M3 Ultra comparison covers the other models.

What changes if you are buying

• The #1 runs on a 24 GB Mac. Qwen 3.6-35B-A3B is tight on 24 GB (Ollama’s default tag is 23 GB) and comfortable on 32 GB, where it runs at community64 tok/s community-measured on a Mac mini M6. The 24 and 32 GB guide covers which build to pick.
• 192 GB and up: MiMo-V2.6-Flash is still the pick. It is slower than estimated, not slow, and it still tops the per-Mac ranking on every Mac from 192 to 512 GB.
• 64 GB: the pick changed to Liquid AI’s LFM2.5-24B-A2B, below.

The index’s pick per RAM tier as of 24 October 2026, quoted on each tier’s representative Mac. The full lists are on best LLM by RAM.

Why 64 GB moved: the per-Mac ranking adds capability to a capped speed bonus (tok/s × 0.12, at most 15 points) for that Mac, so measured speed on the tier’s representative Mac, the Mac mini M5 Pro, matters. On 18 October both models had only estimates there. Now LFM2.5-24B-A2B has two community oMLX runs on an M5 Pro (122.0 and 113.5 tok/s), and Qwen 3.6-35B-A3B has seven measured M5 Pro runs whose lower median is 79.48. That gives LFM2.5 48.7 points to Qwen’s 47.5. The same happens on the 24 GB M5 Pro Macs. LFM2.5 is a reasoning model built for agents and tool calls; for coding, Qwen 3.6-35B-A3B is still the stronger model (Liquid’s own SWE-bench Verified run puts LFM2.5 at 51.0 and Qwen at 66.2 on one harness). Its LFM Open licence allows commercial use only below US$10M in annual revenue.

What would put MiMo back at #1

MiMo-V2.6-Flash needs a displayed M5 Ultra figure of 75 tok/s or more for 19 speed points. That makes 81, a tie with Qwen 3.6-35B-A3B that MiMo wins on name order. In runs:

• Two more eligible runs at 74.5 tok/s or faster. With 61.7 and two such runs, the lower median of three is the middle one, which displays as 75 or more.
• One such run, if the 76.4 run’s hardware is confirmed. A corrected label would make that run eligible, and three runs would then have a middle value of at least 74.5.
• An outright #1 (Score 82) needs a lower median of 77.5 tok/s or more.

Eligible means plain decoding (MTP and drafting off), 5 bits or fewer, a released runtime and a hardware record that matches a Mac Apple sells. Runs of the MOPD checkpoint the catalog lists are the most useful. Nothing guarantees a faster result: more runs could just as well land near 62. MiMo’s drop rests on a single community run on a release-candidate build, next to a sibling run with a contradictory hardware record, and this post will be updated as runs arrive. To add one, see how to submit a run.

The other new M5 Ultra and M5 Pro runs

This update added 92 measured runs to the index. These are the ones on the M5 Ultra and M5 Pro, apart from MiMo’s above and the new models, which have their own posts.

M5 Ultra

Six of these come from Bohemian Silicon, a Czech company that rents out the 256 GB Mac Studio M5 Ultra it tested on (its write-up, 15 October). It ran mlx-lm 0.32 on macOS 27.0.1 with mlx-community 4-bit builds (MXFP4 experts for the gpt-oss models, QAT 4-bit for Gemma 4 31B), a 512-token prompt and 128 generated tokens. The index records them as community runs because the publisher is not a press outlet and has a commercial interest in the machine. Its 8K and 32K-context figures are long-prompt results, so they appear under each row as prose, not as rows.

M5 Pro

Two sources supply most of these. Frandroid’s review of NVIDIA’s RTX Spark laptop (15 October) used a 48 GB MacBook Pro 14 M5 Pro as a comparison, with llama.cpp b11582 on Metal and 4-bit GGUFs whose type it does not name. They are the first press figures for the M5 Pro in the index, so they are labelled published. The developers of the Kapitsa engine ran about ten runtimes on another 48 GB MacBook Pro M5 Pro with their own harness (19 October; community); the index takes rows only from their plain-decode runs, with speculation and multi-token prediction off.

The spread is the lesson. Qwen 3.6-35B-A3B ran from 68.7 to 95.6 tok/s across seven runs on two 48 GB MacBook Pros, depending on the runtime. The index shows the lower median, community79 tok/s on the M5 Pro, and lists every run on the Qwen 3.6-35B-A3B on M5 Pro page.

Also in this update

• Recursive Flint (Apache 2.0, 40B with 2.5B active), a coding-agent model, enters at #2: Recursive Flint on a Mac.
• Nemotron 3.5 Bolt E4B, NVIDIA’s small model for 8–16 GB Macs: Bolt E4B on a Mac.
• MiniMax-M3.1-Flash, a 428B model for 256 GB Mac Studios, community-measured at 34.5 tok/s on an M5 Ultra and 25.3 on an M3 Ultra: MiniMax-M3.1-Flash on a Mac Studio.
• Raon-1 mini preview (KRAFTON) and Spark-X2.6-4B (SparkLLM) join the catalog too. Raon needs a forked or remote-code runtime on a Mac, so it is ranked but never a per-Mac headline pick.
• Two estimator rules. A full chip can no longer be estimated below a measured run on a cut-down version of the same chip (the Flint case), and GPT-oss runs whose source did not state the quant count as MXFP4, the format they ship in. Both are on the methodology page.
• The catalog now holds 116 models and 480 speed rows: 47 published, 146 community and 287 estimated. The changelog lists everything added, and what was left out and why.

Frequently Asked Questions