The 30-Day Recap (TL;DR)
- A tie at #1. MiMo-V2.6-Flash and Qwen 3.6-35B-A3B both score 81. MiMo-V2.6-Flash is listed first on the alphabetical tie-break; it runs at 77 tok/s community-measured on an M5 Ultra but needs a Mac Studio with 192 GB or more. Qwen 3.6-35B-A3B needs 24 GB and is the best pick on a Mac with 32 GB or less.
- GLM-4.7-Flash jumps from #61 to #3 (Score 80). The index had catalogued Zhipu AI’s 30B-A3B mixture-of-experts model as a dense 31B model since January.
- Recursive Flint gets a stock-runtime figure: 129.6 tok/s on a full 40-core M5 Max with llama.cpp (community). Its leaderboard speed is now a measurement, 123 tok/s, and its Score stays 80.
- Fourteen models added, from MiniCPM5-2B (the new 8 GB pick) to Mistral Large 4 (512 GB only) and the server-class Kimi K3.1 Pro.
- Hardware: a 14-inch MacBook Pro with M6 (from $1,999), an iMac with M6 (from $1,499), a PROLED MacBook Pro on the same M5 Pro and M5 Max chips, and the 512 GB M5 Ultra Mac Studio, orderable since 28 October at $22,299. The index also corrected MacBook prices that predated Apple’s June increase and added the MacBook Air M5 it had missed since March.
- Data: 126 measured rows added (100 community, 26 published), so 319 of the index’s 619 speed rows are now measurements; 130 models; 54 Mac configurations.
- Runtimes: llama.cpp 1.0.0 (13 November); DeepSeek V4.1 Flash in upstream llama.cpp; Ollama 0.40.16; mlx-vlm moved to a new GitHub home.
A Tie at #1: Why MiMo-V2.6-Flash Is Back
On 24 October MiMo-V2.6-Flash fell to #7 (Score 78). Its leaderboard speed then rested on one plain-decode run on an M5 Ultra, 61.7 tok/s, and the index had set aside a second, 76.4 tok/s, because it came from a 64-core-GPU M5 Ultra with 256 GB, which the index wrongly said Apple does not sell. Both halves of that have changed.
More runs. The oMLX community board holds many more MiMo-V2.6-Flash runs on 256 GB M5 Ultras with multi-token prediction, DFlash and SpecPrefill all off in each page’s own settings. Most were posted between 24 September and 26 October, and the October updates missed them. The index now holds 15 such sessions, eleven on the 80-core GPU and four on the 64-core GPU, from 61.0 to 89.4 tok/s across the RL and MOPD checkpoints at 4-bit and MXFP4, all community-measured. The MOPD checkpoint the catalog lists ran at 78.7 tok/s on oMLX 0.7.0 (run bq9yj6vv).
One row per measurement session. The board lists every run separately, and a single 26 September session posted 21 of them. Counting each run would let one machine outvote everyone else, so the index enters a session once: its 1K-token run, or its shortest context when it has no 1K run, and the session median when that run was repeated. oMLX .dev and rc builds count as released runtimes, as the reference-row rule has said since October.
The 64-core configuration is sold. Apple’s store offers the Mac Studio with a 30-core-CPU, 64-core-GPU M5 Ultra and 256 GB at $9,499 (checked 14 and 15 November). Both GPU versions have the same 1.2 TB/s of memory bandwidth, so under the index’s rules the 64-core part is not a binned, lower-bandwidth configuration, and its four sessions count.
The result. The lower median of the 15 sessions is 76.8 tok/s (a 64-core run on oMLX 0.7.3, 14 November), shown as 77, community-measured. That is worth 19 speed points, so MiMo-V2.6-Flash scores 49 (capability) + 19 (speed) + 3 (memory) + 10 (MIT licence) = 81. Qwen 3.6-35B-A3B scores 38 + 25 + 10 + 8 = 81. Ties break by name, so MiMo-V2.6-Flash is listed first, and every surface that names the #1 says it is a tie. Counting only the 80-core sessions would give 78 tok/s and Score 82, #1 outright; the index has no grounds to drop the 64-core sessions, so it does not.
The figure is physically plausible. With 15B active parameters at about 4.26 bits each, MiMo-V2.6-Flash reads roughly 8 GB per token, which caps an M5 Ultra near 154 tok/s; 77 is half of that, and about 1.3 times the 55–59 tok/s community runs on M3 Ultras.
What it means for buyers: nothing changes below 192 GB. MiMo-V2.6-Flash starts at a 192 GB Mac Studio (its smallest GGUF is 126 GB; 4-bit MLX builds are about 172 GB) and is the pick for 192–512 GB. On a Mac with 32 GB or less the pick is Qwen 3.6-35B-A3B, whose own M5 Max figure moved from 118 tok/s published to 130 community-measured, the lower median of eight runs (118–153).
GLM-4.7-Flash: a Correction, Not a Breakthrough
The month’s biggest move is a fix. Zhipu AI released GLM-4.7-Flash under MIT on 19 January; its model card calls it a 30B-A3B mixture-of-experts model, and its configuration (glm4_moe_lite: 64 routed experts, 4 active per token, plus a shared expert) agrees. The index had listed it as a dense 31B model, so it costed every token as a read of all 31B weights: about 26 tok/s estimated on an M5 Max, Score 61, #61.
A community oMLX run exposed the error: 46.6 tok/s on a 32 GB M6 (oQ4e, multi-token prediction off). A dense 31B model at 4-bit cannot pass about 10 tok/s on the M6’s 170 GB/s. Costed as a MoE with about 3B active, GLM-4.7-Flash estimates at 107 tok/s on an M5 Max, takes the full 25 speed points and scores 80: #3, ahead of Recursive Flint on the alphabetical tie-break. Its 18 GB minimum fits a 24 GB Mac. Two caveats: its M5 Max figure is still an estimate, and its capability score (35) is below Qwen 3.6-35B-A3B’s (38). The rank comes from speed, a small memory footprint and the MIT licence.
Recursive Flint’s First Stock M5 Max Run
Since its 21 October release, Recursive Flint’s M5 Max figure had been an estimate held at about 131 tok/s by a community run on the binned 36 GB M5 Max (460 GB/s), because no qualifying run on a full 40-core M5 Max existed. Now one does: 129.6 tok/s with llama.cpp b11871 and Recursive’s Q4_1 GGUF on a 40-core MacBook Pro 16-inch, from a published benchmark log (community-measured, 26 October). Four oMLX runs on full M5 Max chips joined it (77.4–132.5 tok/s, community), so the leaderboard now shows the lower median of the five: 123 tok/s, community-measured (oMLX, 4-bit). The Score stays 80, because speed points cap at 100 tok/s.
Two kinds of run are shown but not counted: two runs of an unidentified “mixed 3/4-bit” build (163.5 and 147.2 tok/s) with no public repository or stated bits per weight, and custom inference engines in the same log that report roughly twice the llama.cpp figure. The binned 36 GB run (131) now sits above the full chip’s lower median (123); the index shows both, as two real measurements that disagree. On a 24 GB Mac mini M6 Flint is estimated at about 71 tok/s, which makes it the new 24 GB pick.
Open-Source Top 10 (November 2026)
| # | Model | Score | License | Min RAM | Speed (M5 Max unless noted) |
|---|---|---|---|---|---|
| 1 | MiMo-V2.6-Flash | 81 | MIT | 170 GB | 77 tok/s community (M5 Ultra) |
| 2 | Qwen 3.6-35B-A3B | 81 | Apache 2.0 | 24 GB | 130 tok/s community |
| 3 | GLM-4.7-Flash | 80 | MIT | 18 GB | 107 tok/s est. |
| 4 | Recursive Flint | 80 | Apache 2.0 | 20 GB | 123 tok/s community |
| 5 | KAT-Coder-V2.5 | 79 | Apache 2.0 | 24 GB | 89 tok/s community |
| 6 | Nemotron 3.5 Lightning | 79 | OpenMDW | 18 GB | 117 tok/s community |
| 7 | North Mini Code 1.0 | 79 | Apache 2.0 | 21 GB | 120 tok/s community |
| 8 | Gemma 4 26B-A4B | 78 | Apache 2.0 | 18 GB | 107 tok/s published |
| 9 | MiniCPM-V 4.7 35B-A3B | 78 | Apache 2.0 | 23 GB | 100 tok/s community |
| 10 | Raon-1 mini preview | 77 | Apache 2.0 | 18 GB | 123 tok/s est. |
Nine of the ten are mixture-of-experts models of 26–40B total parameters with 2.5–4.3B active, and all nine fit a 24 GB Mac; MiMo-V2.6-Flash is the exception. Eight of the ten speeds are measurements; GLM-4.7-Flash’s and Raon-1 mini preview’s are estimates. New in the Top 10 since 24 October: GLM-4.7-Flash. Out: LFM2.5-24B-A2B (#11, Score 76), still the 64 GB pick. The highest capability score a 24 GB Mac can hold, 46 of 50, is now shared by Qwen3.8-27B (#14, Score 72; 31 tok/s community-measured on an M5 Max) and the new Nanbeige4.3-33B. Every row, with its provenance mark, is on the leaderboard.
The Best Pick for Each RAM Tier
Each tier’s pick is the best-ranked model that fits the tier, reaches 15 tok/s on the tier’s representative Mac and runs on a standard Mac runtime. The 24 and 32 GB tiers are now listed separately.
| Memory | Pick | Speed on the reference Mac | Score | Needs |
|---|---|---|---|---|
| 8 GB | MiniCPM5-2B (new; was Nanbeige4.2-3B) | ~39 tok/s estimated, MacBook Air M1 8 GB | 69 | 3 GB |
| 16 GB | Gemma 4 12B | ~17 tok/s estimated, Mac mini M6 16 GB | 66 | 9 GB |
| 24 GB | Recursive Flint (was Qwen 3.6-35B-A3B) | ~71 tok/s estimated, Mac mini M6 24 GB | 80 | 20 GB |
| 32 GB | Qwen 3.6-35B-A3B | 59 tok/s community-measured, Mac mini M6 32 GB | 81 | 24 GB |
| 64 GB | LFM2.5-24B-A2B | 114 tok/s community-measured, Mac mini M5 Pro 64 GB | 76 | 17 GB |
| 128 GB | Qwen 3.6-35B-A3B (was Qwen3.8-Flash-Next) | 130 tok/s community-measured, Mac Studio M5 Max 128 GB | 81 | 24 GB |
| 192–256 GB | MiMo-V2.6-Flash | 77 tok/s community-measured, Mac Studio M5 Ultra 256 GB | 81 | 170 GB |
| 512 GB | MiMo-V2.6-Flash | 77 tok/s community-measured on the M5 Ultra (runs on 256 GB machines) | 81 | 170 GB |
Two picks are close calls on the published suitability formula (capability score plus 0.12 points per tok/s, capped at 15). At 24 GB, Recursive Flint scores 45.5 against Qwen 3.6-35B-A3B’s 45.1; a new 58.7 tok/s community run of Qwen on a 32 GB M6 lowered its M6 figure from 64 to 59 and tipped the balance. At 128 GB, Qwen 3.6-35B-A3B (53.0, its speed bonus capped) edges Qwen3.8-Flash-Next (52.6, 47 tok/s community-measured on an M5 Max). The 128 GB pick is the model that runs best there, not the largest one that fits: the most capable model a 128 GB Mac can hold is GLM-5.3-Flash, at about 27 tok/s estimated. Exact figures for all 54 configurations are on the by-RAM page and the Best-by-Mac pages.
Fourteen New Models, by the Mac They Need
Six of the fourteen were released on or after 24 October. Six more are older releases the October changelog had queued as “seen, not yet evaluated”, and two (K2-Horizon-3.7B and Nemotron 3 Ultra) are catalog gaps the index had missed. Each was checked against its Hugging Face card.
8 GB Macs and up
- MiniCPM5-2B (OpenBMB, Apache 2.0, 6 September; 2.5B dense, 128K context). Artificial Analysis names it and K2-Horizon-3.7B the two highest-scoring open models of 4B parameters or fewer. Q4_K_M is 1.6 GB; stock llama.cpp and the official MLX build run it, and SparkLLM’s model-card chart lists 147 tok/s on an M5 Pro (published; mlx-lm 0.32.2, 4-bit, 4K context). It is the new 8 GB pick, at about 39 tok/s estimated on an 8 GB MacBook Air M1. Score 69.
- K2-Horizon-3.7B (MBZUAI’s Institute of Foundation Models, Apache 2.0, 1 September; 3.7B core parameters, 5.1B with its embeddings, 512K context, training data and code published). It has the higher capability score of the two small leaders (22 against 21) and ranks higher (#19, Score 70), but its 3.2 GB Q4_K_M file reads about half as fast on an M1’s 68 GB/s (about 18 tok/s estimated), so MiniCPM5-2B stays the 8 GB pick. Upstream llama.cpp supports it from b11454; mlx-lm does not yet (pull request #1948).
- LFM3-1.7B (Liquid AI, LFM Open License, 27 October; a 1.7B “tri-hybrid” model, about 1 GB at 4-bit, reasoning on by default). Ollama added it as
lfm3in 0.40.12; on MLX it runs through mlx-vlm 0.7.11 or later. Community oMLX runs measured 116.7 tok/s on a base M5 and 92.5 on an M4 (4-bit). Score 63. - Ling-3.0-tiny (Ant Group, MIT, 10 August; 7.9B mixture-of-experts with 1.3B active). A community oMLX run measured 174 tok/s on an M5 Max (oQ4e, 4K context; one session counts, and a second, erratic session is shown but not used). Its only capability evidence is vendor-quoted, so the index holds its capability score at 15, below Gemma 4 E4B’s. It is the fastest option on an 8 GB Mac, about 87 tok/s estimated on a MacBook Air M1. Score 65.
24–32 GB Macs
- Nanbeige4.3-33B (Nanbeige, Apache 2.0, 24 October; dense 33.1B, 512K context). Artificial Analysis names it and Qwen3.8-27B the two highest-scoring small open models (4–40B). Vendor-run card: Terminal-Bench 2.1 80.9, SWE-Bench Pro 60.3 (mini-swe-agent, not Scale’s SEAL harness), GPQA Diamond 89.5. Q4_K_M is 20.8 GB, so it targets a 24 GB Mac; stock llama.cpp runs it. Community oMLX runs measured 24.5 and 27.4 tok/s on an M5 Max at 4-bit. As a dense model it is slow on an M6, about 7 tok/s estimated on a 24 GB Mac mini. Score 70 (#22). No Ollama tag yet.
48–64 GB Macs
- Ripple-65B Preview (Ricursive Intelligence, Apache 2.0, 4 November; a dense 64.8B model trained from scratch, 1M context). Ricursive Intelligence is not related to Recursive, the maker of Flint. Self-reported card: SWE-bench Verified 76.5 (OpenCode harness), Terminal-Bench 2.1 80.5. Q3_K_M is 31 GB and Q4_K_M 38.8 GB, so plan on 48–64 GB. Upstream llama.cpp added it on 5 November (pull request #31692) and Ollama 0.40.16 ships it as
ripple-preview; its MLX builds still bundle their own model file, so the community MLX runs (17.2 tok/s on an M3 Ultra, 23.3 on an M5 Ultra, both 4-bit) are shown but do not anchor estimates. About 13 tok/s estimated on an M5 Max; Score 59. - Ling-3.0-flash-VL (Ant Group, MIT, 4 September; a 124B vision-language MoE with 5.5B active; text, image and video in). Artificial Analysis names it and Nemotron 3.5 Super VL the two highest-scoring medium open models (40–150B). IQ3_XXS is 54 GB, so it fits a 64 GB Mac; llama.cpp b11157 and later run it, MLX does not yet. About 53 tok/s estimated on an M5 Max; Score 70 (#20).
96–128 GB Macs
- Step-3.7-Flash (StepFun, Apache 2.0, 23 May; 198B MoE with 11B active, image input). It supersedes Step-3.5-Flash, which stays listed. The 2-bit oQ2e MLX build is 63 GB, so it needs a Mac with more than 64 GB. Community oMLX runs measured 48.2 tok/s (oQ2e) and 51.3 (oQ3e) on an M5 Max, and 68.5 at 4-bit on a 64-core M5 Ultra. Score 64.
- dots3-note preview (Xiaohongshu, or RedNote; Apache 2.0, 14 August; 280B MoE with 16B active; text, image, video and audio in, 512K context). Vendor chart: SWE-bench Pro 61.0, Terminal-Bench 2.1 75.1 (harness not stated). bartowski’s IQ3_XXS is 117.5 GB, so it needs a 128 GB Mac; upstream llama.cpp runs it. No Mac speed on a released build has been published; about 28 tok/s estimated on an M5 Max. Score 62.
256 GB Mac Studios
- Inkling 2 Small (Thinking Machines, Apache 2.0, weights 12 November; 594B MoE with 17B active, trained natively in NVFP4; text and image in, 1M context). Card: Terminal-Bench 2.1 81.3, GPQA Diamond 93.0. Unsloth’s UD-IQ3_XXS build is 226 GB, a tight fit on a 256 GB Mac Studio. At this edition’s cutoff only oMLX’s main branch ran it (pull request #5229, merged 13 November, after 0.8.0rc1); the oMLX developer reports 23.3 tok/s on an M3 Ultra at a 4K context with multi-token prediction off (community, prose only, as no released runtime ran it), so it carries a runtime caveat and cannot be a Best-by-Mac headline pick. About 36 tok/s estimated on an M5 Ultra; Score 68.
- Nemotron 3 Ultra 550B-A55B (NVIDIA, OpenMDW, 3 June; a hybrid Mamba-Transformer MoE with 55B active, 1M context). A catalog gap: released in June and never listed until now. Card: SWE-bench Verified 70.7, GPQA Diamond 87.0. Unsloth’s UD-IQ3_XXS build is 223 GB, tight on 256 GB; llama.cpp runs its architecture, but no Mac speed has been published. About 12 tok/s estimated on an M5 Ultra; Score 56.
512 GB Mac Studios
- Mistral Large 4 (Mistral, weights 29 October; about 1.05T total with 49B active, 1M context, image input). It ships under a new Mistral Ecosystem Open-Weight License: an MIT-style grant with one condition, that a company running a Model-as-a-Service business needs a separate commercial licence from Mistral. Internal use and end-user products that embed the model are exempt. The index gives it 5 of 10 licence points, the class of the other licences that need a separate agreement for API hosting. Q2_K is 381 GB, so the 512 GB Mac Studio is the only single Mac that holds it. Hugging Face’s launch post reports 15 tok/s generation on a 512 GB M3 Ultra at Q2_K with llama.cpp (published); the index estimates about 21 tok/s on an M5 Ultra. llama.cpp supports it from b11998 (pull request #31376) and mlx-vlm from 0.7.12; stock mlx-lm does not yet (#2018). Score 58. What it means for a buyer: Mistral Large 4 on a 512 GB Mac Studio.
- Hy4 preview (Tencent, Apache 2.0, 27 August; 770B MoE with 49B active, 1M context; its card calls it an early version with known issues). Vendor appendix: SWE-bench Pro 65.7, GPQA Diamond 92.3. Its smallest GGUF of about 3 bits per weight (6block’s Q2_K) is 289.5 GB, so it needs 512 GB. llama.cpp has its architecture, validated only on CUDA, and no Mac run has been published, so it carries a runtime caveat. About 14 tok/s estimated on an M5 Ultra; Score 63.
Server-class
- Kimi K3.1 Pro (Moonshot AI, weights 2 November, two days before Moonshot’s 4 November promise; a 2.8T MoE with 104B active, text and image in, 1M context, under a custom licence with Kimi K3’s clauses). Artificial Analysis names it and DeepSeek V4.1 Flash 1017 the two highest-scoring large open models. It ships as QAT MXFP4 at 1.56 TB, and even the 1-bit GGUF (593 GB) exceeds a 512 GB Mac, so it is listed without a Mac figure (Score 58). llama.cpp 1.0.0 lists its architecture among its additions. Kimi K3 stays listed, marked as superseded.
Considered and Not Added
Fine-tunes and derivatives. A derivative is a checkpoint built on a catalog model: continued training, a supervised or RL fine-tune, a merge, an abliteration or a domain specialisation, or a checkpoint whose architecture is identical to one. The index lists a derivative only when all four of these hold: an accountable organisation publishes it with a licence file; an independent evaluation, or a vendor table against its own base with the harness stated, shows a gain worth at least two capability points; it has a downloadable Mac build on a stock runtime; and it is more than a safety-removal, persona or style change. A lab’s own updated checkpoint replaces its entry instead, as DeepSeek V4.1 Flash 1017 and Ling-3.1-flash-1019 did. None of this month’s derivatives met all four: Arete-27B (a supervised fine-tune of Qwen3.8-27B), Sage-A1.5 (a Qwen 3.6-35B-A3B fine-tune), Atria Dawn (built on GLM-5.2; no licence file, no Mac build of the final weights), Ornith-1.5 (licence file still missing), Hydrogen-1.0 and Nex-N2.6 (GLM-5.3 and GLM-5.3-Flash fine-tunes), Intern-S2-397B (Qwen 3.5 397B-A17B’s configuration with an extended vocabulary) and Ling-3.0-flash-Fin (finance continued training).
No practical Mac path. LongCat-2.5 (1.6T, MIT) is published for NPU deployment and is 1.34 TB even in its smallest format; LongCat-2.0 is superseded and server-class; ERNIE-5.5-Flash-Base is a base checkpoint only, under a licence that excludes the EU, the UK and South Korea; Kimi K3.1 Highspeed is API-only.
Size class and niche. MiniCPM5-1B and Spark-X2.6-1.7B (the index lists their larger siblings); Raon-1 nano (KRAFTON’s fork runtimes only; its upstream llama.cpp pull request, #32200, opened on 13 November); SenseNova-Loop-9B-Preview (CUDA only); and draft, translation, image, OCR and world models.
No weights. Qwen 4: Alibaba’s Qwen/Qwen4-27B repository is not public, and the community “Qwen4-27B” repositories on Hugging Face are empty placeholders. Nothing named GLM-5.5 or GLM-6 is on Hugging Face. Every addition and exclusion is logged in the changelog.
Hardware: M6 MacBook Pro, iMac and a $22,299 Studio
Apple announced three Macs on 27 and 28 October; pre-orders opened on 28 October, and all three shipped on 6 November. None has a new Pro, Max or Ultra chip, and Apple published no generation speeds: its M6 footnotes cite only prompt processing (time to first token up to 1.3 times faster than the M5 MacBook Pro), which says nothing about how fast a model writes.
| Mac | Chip and bandwidth | Memory | Apple price | Index pick |
|---|---|---|---|---|
| MacBook Pro 14-inch (M6), new | M6; 153 GB/s at 16 GB, 170 at 24–32 GB | 16, 24, 32 GB | from $1,999 (1 TB) | 16 GB: Gemma 4 12B; 24 GB: Recursive Flint; 32 GB: Qwen 3.6-35B-A3B |
| iMac (M6), new | M6, 10- or 12-core GPU; “up to 170 GB/s” | 16, 24, 32 GB | from $1,499 | 16, 24 and 32 GB: as the Mac mini M6 |
| MacBook Pro with PROLED XDR, new | M5 Pro 307 GB/s; M5 Max 460 or 614 GB/s | 24–128 GB | from $2,899 (M5 Pro), $4,499 (M5 Max) | as the mini-LED MacBook Pro |
| MacBook Air (M5), newly listed | M5; 153 GB/s | 16, 24, 32 GB | from $1,299 | 16 GB: Gemma 4 12B; 24 and 32 GB: Qwen 3.6-35B-A3B |
| Mac Studio M5 Ultra, 512 GB, now orderable | 36-core CPU, 80-core GPU only; 1,228 GB/s | 512 GB | $22,299 (1 TB) | MiMo-V2.6-Flash |
The M6 MacBook Pro has the Mac mini M6’s chip and bandwidth, against 153 GB/s for the M5 MacBook Pro it replaces, so the Mac mini M6 figures carry over: Qwen 3.6-35B-A3B at 59 tok/s community-measured on a 32 GB M6, dense 27B models at about 10. The iMac’s spec page gives only “up to 170GB/s”, with no split by memory; the index assumes 153 GB/s at 16 GB, as Apple specifies for the other M6 Macs. The PROLED MacBook Pro uses the same M5 Pro and M5 Max chips as the mini-LED model, which stays on sale, so it runs models at the same speed; the $400 premium buys the tandem-OLED touch display. The MacBook Air M5 has been on sale since March; the index had skipped it and now has three pages for it.
The 512 GB Mac Studio has been orderable since 28 October, only with the 36-core CPU and 80-core GPU M5 Ultra: $22,299 with 1 TB (checked on Apple’s store on 15 November), $11,500 more than the same chip with 256 GB ($10,799). The 64-core 256 GB model is $9,499. Apple quoted 4–6 business days for delivery on 14 November. What 512 GB buys: 14 more of the index’s models fit than at 256 GB (122 against 108), among them Mistral Large 4, Muse Spark Open, DeepSeek V4.1 Flash, Ling-3.1-flash and Kimi K2.6, but only in 2- to 4-bit-class builds and at about 14–31 tok/s estimated on an M5 Ultra, and no plain-decode run on a 512 GB M5 Ultra had been published by 14 November. M5 Ultra vs M3 Ultra vs M5 Max compares the three Studio chips.
The June price rise. Apple raised Mac prices on 25 June, citing the cost of memory (MacRumors): the 14-inch MacBook Pro with M5 Pro went from $2,199 to $2,499, and the M5 Max model now starts at $4,099. The index’s MacBook prices predated the increase, one by $2,800 (it listed the 128 GB M5 Max MacBook Pro at $3,899; Apple charges $6,699), and several build-to-order Mac Studio and Mac mini prices were approximations. Every current configuration now carries Apple’s list price, checked on 14 November.
Previous generation. The Mac mini M4 and M4 Pro, MacBook Air M4, and the Mac Studio with M4 Max or M3 Ultra are no longer sold new (Apple’s store redirects their configuration pages). The index now labels them previous generation. New Best-by-Mac pages this month: the three M6 MacBook Pros, three M6 iMacs and three M5 MacBook Airs, plus the 48 GB MacBook Pro M5 Pro, the 48 GB Mac Studio M5 Max (at $3,099, the cheapest Mac with 614 GB/s) and the 48 GB Mac mini M4 Pro (used).
Measured This Month
The index added 126 measured rows, 100 community runs and 26 published press or vendor figures, so 319 of its 619 speed rows are now measurements; everything else is the formula’s estimate, labelled as one. The best-documented sources:
- Vše Smart, a Czech Apple site, published two reviews with their setups stated. On a MacBook Pro M5 Max (40-core GPU, 128 GB; llama.cpp b11669 and mlx-lm 0.32.2), GPT-oss 120B ran at 110 tok/s on llama.cpp and 94 on mlx-lm (published, MXFP4). On a Mac Studio M5 Ultra (80-core GPU, 256 GB), Qwen3-235B-A22B ran at 46 tok/s on MLX (published, 4-bit). With a 92.0 tok/s oMLX run (community), GPT-oss 120B’s M5 Max figure is now 92 tok/s community-measured, the lower median of four runs, and its Score rises from 64 to 68.
- Bohemian Silicon, a Mac rental firm, measured Qwen 3.5 397B-A17B at 61.6 tok/s on a 256 GB M5 Ultra (community; mlx-lm 0.32.2, 512-token prompt, 224 GB peak memory), against the index’s estimate of 36. Its Score rises from 58 to 65.
- The published M5 Max benchmark log behind the Flint figure also gave stock-runtime rows for Qwen 3.6-35B-A3B (122.0 tok/s mlx-lm, 139.6 LM Studio) and Qwen3.8-27B (31.0 and 32.9), all community-measured.
- M6 coverage from community runs on 32 GB M6 Macs, at about 4 bits: Qwen 3.6-35B-A3B 58.7 tok/s, Nemotron 3.5 Lightning 60.3, Gemma 4 26B-A4B 49.4, GLM-4.7-Flash 46.6, GPT-oss 20B 65.6, Gemma 4 12B 19.4 and Qwen3.8-27B 9.7–11.5.
- New models: Nanbeige4.3-33B on four chips, Step-3.7-Flash on four, LFM3-1.7B on four, Ling-3.0-tiny and MiniCPM5-2B on two each, and Hugging Face’s Mistral Large 4 figure.
Still missing: no review of the M6 MacBook Pro or iMac has published a fully specified generation figure, and nobody has posted a plain-decode run on a 512 GB M5 Ultra.
Runtimes: llama.cpp 1.0, Ollama, MLX
Update, 19 November 2026. Two releases since this edition: oMLX 0.8.0 (18 November) is the current stable release and includes Inkling 2 Small support (pull request #5229), and mlx-vlm 0.8.0 shipped on 16 November. The figures below are as published on 16 November.
- llama.cpp 1.0.0 (13 November) is the project’s first 1.x release and corresponds to build b12359; Homebrew’s llama.cpp formula moved to it. It removes the deprecated API and adds FP8 KV caches and a host-memory cache for MoE experts. Homebrew users now get the fixes for Nemotron 3.5 Bolt (b11755), MiniMax-M3.1-Flash (b11861) and DeepSeek V4.1 Flash (b11870).
- DeepSeek V4.1 Flash in upstream llama.cpp. Pull request #31086 merged on 25 October and shipped in b11870; the older pull requests the index had been tracking (#28696 and #30817) were closed. No qualifying plain-decode Metal figure exists yet: AI Runner’s llama.cpp runs on a 256 GB M5 Ultra (about 42 tok/s at the 128K position with Unsloth’s patched build, community) are 2-bit, long-context figures, which the index keeps in prose. Stock mlx-lm still refuses its Engram layers, and Ollama offers it in the cloud only. The index dropped its runtime caveat; on a 512 GB Mac Studio it is estimated at 29 tok/s (M5 Ultra, oQ4e). Score 70.
- Ollama 0.40.12 to 0.40.16 (27 October to 9 November) added LFM3 (
lfm3), MiMo-V2.6-Distill-Qwen-9B (mimo-v2.6) and Ripple-65B Preview (ripple-preview); in 0.40.16, Qwen3.8-Flash-Next uses about 27 GiB less memory on Apple silicon. Nemotron 3.5 Bolt 74B on Apple Silicon (#19229) is still an open pull request. - mlx-vlm moved to github.com/Neywa-Labs/mlx-vlm (the old address redirects). Version 0.7.11 (28 October) added MiniMax-M3.1-Flash vision; 0.7.12 (4 November) added Raon-1 mini and Mistral Large 4.
- oMLX 0.7.3 (31 October) is the current stable release and serves Raon-1 mini without remote code; Inkling 2 Small support (pull request #5229) merged into its main branch on 13 November, after 0.8.0rc1 (12 November), and is not in a release yet.
- mlx-lm is still 0.32.2 (16 October) and MLX 0.32.4 (14 October). Merged but unreleased: Recursive Flint support (#1998) and original-checkpoint loading for MiniMax-M3.1 (#2003) in mlx-lm, and Neural Accelerator matrix kernels in MLX (#4781), which speed up prompt processing and batched work, not single-stream generation. Open: Mistral Large 4 (#2018), Inkling 2 (#2061), LFM3 (#2016), Raon-1 mini (#2015) and K2-Horizon (#1948).
- Raon-1 mini preview lost its runtime caveat: it runs in oMLX 0.7.3 and mlx-vlm 0.7.12, and community oMLX runs measured 114.4 and 129.1 tok/s on a 40-core M4 Max (oQ4e and NVFP4). KRAFTON opened its own upstream llama.cpp pull request (#32176) on 13 November. Anchored on those runs, its M5 Max estimate rose from 99 to 123 tok/s; Score 77, #10.
- Ling-3.1-flash now points at inclusionAI’s 27 October “1019” checkpoint, which has public MLX builds; oMLX’s 0.7.3rc1 release notes report 18.8 tok/s on a 512 GB M3 Ultra with multi-token prediction off (community, 4K context).
What the New Runs Say About the Estimator
For mixture-of-experts models above 40B total parameters, the index multiplies its bandwidth estimate by 0.42, a factor fitted in September to DeepSeek V4 Flash. This month’s large-Qwen runs came in well above it: Qwen 3.5 397B-A17B at 61.6 tok/s against an estimate of 36 (1.7 times), and Qwen3-235B-A22B at 46 against 29 (1.6 times), both on M5 Ultras. Across the 18 large MoEs that now have qualifying runs, measured speed lands at 0.26 to 0.88 of the unscaled formula, median 0.57; the per-model list is on the methodology page.
The index keeps 0.42 anyway, for two reasons. It is meant as a conservative default, and three of the 18 measured below it (DeepSeek V4.1 Flash, Qwen3.8-Flash-Next and Mistral Large 4), so a higher factor would overstate some models. And it now decides little: a model with its own measured rows is scaled from those rows, not from the factor, so 0.42 sets the figure only for large MoEs nobody has measured. Both Qwen models are now anchored on their own runs; Qwen3-235B-A22B’s M5 Max estimate rose from 16 to 26 tok/s. A separate factor for Qwen-family MoEs would need a held-out test before the index adopts one. The small-MoE fit from October and the dense formula are unchanged.
Corrections
- The 64-core M5 Ultra with 256 GB is sold. The 24 October post on Qwen 3.6-35B-A3B’s return to #1 and the dataset note on MiMo-V2.6-Flash’s 76.4 tok/s run said Apple does not offer this configuration, and the run was set aside on that basis. Apple’s store sells it at $9,499. The run now counts.
- GLM-4.7-Flash had been catalogued as a dense 31B model since January; it is a 30B-A3B mixture-of-experts model.
- Mac prices. MacBook prices predated Apple’s June increase, and several build-to-order Mac Studio and Mac mini prices were approximations; every current configuration now carries Apple’s list price.
- Runtime notes. Raon-1 mini preview runs in released MLX tools; DeepSeek V4.1 Flash runs in upstream llama.cpp; Ling-3.1-flash has had public MLX builds since 27 October. Those caveats were removed or rewritten. In the other direction, Inkling and Inkling-Small gained a caveat: their llama.cpp support is still an unmerged pull request (#25731), and stock mlx-lm cannot load them.
- Missing and retired Macs. The MacBook Air M5, on sale since March, was missing, and Macs Apple no longer sells new were listed as current.
What to Watch in December
Nothing here enters the index until weights or measurements exist on a primary source.
- Watch Qwen 4. Still no weights: the Qwen organisation’s newest Hugging Face uploads are image models, and the “Qwen4-27B” repositories elsewhere on the Hub are empty placeholders. llama.cpp’s “qwen4exp” work is Qwen3.8-Flash-Next’s architecture.
- Watch The first plain-decode runs on a 512 GB M5 Ultra: Mistral Large 4 at 3 bits or more, GLM-5.3, Muse Spark Open, and DeepSeek V4.1 Flash on llama.cpp’s Metal backend.
- Watch Inkling 2 Small on stock runtimes. llama.cpp merged its support at 23:51 UTC on 14 November (pull request #32217), after this edition’s data cutoff; it first ships in b12406. mlx-lm support (#2061) is open. A first Mac speed on a released build would let the index lift its caveat.
- Watch MiniMax-M3.1-Flash in llama.cpp. A new pull request for its sparse-attention layers, #32301, replaced #31192 and merged on 15 November (b12420), also after the cutoff.
- Watch Open pull requests: Mistral Large 4 in mlx-lm (#2018); Raon-1 mini in llama.cpp (#32176); LFM3 (#2016), Step-3.7-Flash (#1325), K2-Horizon (#1948) and Kolibri-1 (#1945) in mlx-lm; Nemotron 3.5 Bolt 74B in Ollama (#19229); a Metal validation of Hy4 preview.
- Watch Releases: an mlx-lm release with the merged Flint and MiniMax-M3.1 changes, and an MLX release with the Neural Accelerator kernels (#4781).
- Watch ERNIE 5.5 Flash, if Baidu posts the post-trained model; the base checkpoint’s licence excludes the EU, the UK and South Korea.
- Watch Apple: any M6 Pro, M6 Max or M6 Ultra, and further price rises; Apple tied June’s increase to memory costs, and Micron expects the memory shortage to last through 2027.
- Announced NVIDIA’s acquisition of Hugging Face, agreed on 3 September and expected to close in the first half of 2027.
- Watch Index rules: a rule for runs superseded by a later release of the same runtime, applied to every model at once, and a review of the weakest-provenance rows.
Frequently Asked Questions
What is the best open-source local LLM for a Mac in November 2026?
According to the LLMCheck index (14 November 2026), MiMo-V2.6-Flash and Qwen 3.6-35B-A3B tie for #1 at LLMCheck Score 81, and MiMo-V2.6-Flash is listed first on the alphabetical tie-break. MiMo-V2.6-Flash runs at 77 tok/s community-measured on an M5 Ultra but needs a Mac Studio with 192 GB or more. Qwen 3.6-35B-A3B needs only 24 GB and is the best pick on a Mac with 32 GB or less, at 59 tok/s community-measured on a 32 GB Mac mini M6.
What is the best local LLM for a Mac with 16, 24 or 32 GB?
The index's picks are Gemma 4 12B for 16 GB (about 17 tok/s estimated on a 16 GB Mac mini M6), Recursive Flint for 24 GB (about 71 tok/s estimated on a 24 GB Mac mini M6) and Qwen 3.6-35B-A3B for 32 GB (59 tok/s community-measured on a 32 GB Mac mini M6). The 24 GB call is close: Qwen 3.6-35B-A3B also fits and trails Flint by less than half a point on the index's suitability formula.
What is the best local LLM for an 8 GB Mac?
MiniCPM5-2B, new this month: a 2.5B Apache 2.0 model that needs about 3 GB and runs at about 39 tok/s estimated on an 8 GB MacBook Air M1. K2-Horizon-3.7B has a slightly higher capability score but runs about half as fast there, and Ling-3.0-tiny is the fastest option, at about 87 tok/s estimated.
Why is GLM-4.7-Flash suddenly #3?
Because the index fixed an error. GLM-4.7-Flash is a 30B-A3B mixture-of-experts model, according to its Hugging Face card and configuration, but the catalog had listed it as a dense 31B model since January, which cut its speed estimate about fourfold. Costed correctly, it is estimated at 107 tok/s on an M5 Max and scores 80. A community run measured 46.6 tok/s on a 32 GB M6.
Can a Mac run Mistral Large 4?
Only a 512 GB Mac Studio. Its Q2_K build is 381 GB. Hugging Face's launch post reports 15 tok/s generation on a 512 GB M3 Ultra with llama.cpp (published), and the index estimates about 21 tok/s on a 512 GB M5 Ultra. llama.cpp (b11998 and later) and mlx-vlm 0.7.12 support it; stock mlx-lm does not yet. Its new Mistral Ecosystem Open-Weight License requires a separate agreement only for Model-as-a-Service businesses.
Is the 512 GB Mac Studio M5 Ultra worth $22,299?
Only if you need models above 256 GB. Apple has sold it since 28 October, with the 80-core GPU only, for $11,500 more than the 256 GB model with the same chip. It fits 14 more of the index's models than 256 GB, among them Mistral Large 4, Muse Spark Open and DeepSeek V4.1 Flash, but those run at about 14-31 tok/s estimated, and no plain-decode run on a 512 GB M5 Ultra had been published by 14 November. The index's pick for 192-512 GB, MiMo-V2.6-Flash, fits from 192 GB.
Is the M6 MacBook Pro good for local LLMs?
For models that fit in 32 GB, yes, especially small mixture-of-experts models. It has the Mac mini M6's chip, with 153 GB/s at 16 GB and 170 GB/s at 24 or 32 GB, from $1,999. Qwen 3.6-35B-A3B ran at 59 tok/s community-measured on a 32 GB M6, while dense 27B models manage about 10. For larger models, the M5 Pro (307 GB/s) and M5 Max (614 GB/s) MacBook Pros remain the faster choice.