State of Open-Source Local LLMs — September 2026: Two Frontier MoEs Reach the Mac Studio

The 30-Day Recap (TL;DR)

• Qwen3.8-27B is #1. Apache 2.0, 27.8B dense, native image and video input, a 262K context. August said the top of the leaderboard would change overnight if it shipped open. It did.
• GLM-5.3-Flash is real, and it was “Ox Alpha.” Zhipu confirmed on 26 August that the anonymous model people had been testing was its own 320B-A18B, then put MIT weights on Hugging Face the same day. A 3-bit quant is 120 GB.
• Qwen3.8-Flash-Next previews Qwen4. 125B total, 6B active, image and video input, SWE-Bench Pro 62.5 (vendor scaffold). Its 4-bit MLX conversion is 112 GB, which makes it a 128 GB-tier model — and the new top pick on every 128 GB-and-up Mac page in the index.
• The GLM-5.3 flagship is open too — but not under MIT. Weights went public on 25 August, the day of our last data cut, under a custom GLM-5.3 License that reads like MIT for anyone short of a US$10 billion model-as-a-service business. 753B, ~40B active, a 256 GB Studio job at 2-bit.
• DeepSeek opened V4 Flash Vision. Ten days API-only, then MIT weights on 31 August. Experimental, and the first V4 with image input.
• Muse Spark 1.3 shipped closed. API and Muse Code only. The open-weights promise is still for version 1.2, still without a date.
• The index recalibrated its MoE speed estimates. One published figure — DeepSeek V4 Flash on M5 Ultra — moved from ~80 to ~47 tok/s. Details in the calibration note.

Qwen3.8-27B Takes #1 — as Predicted

Alibaba published Qwen3.8-27B on 14 August under Apache 2.0, and it now sits at the top of the leaderboard with a LLMCheck Score of 71, one point ahead of DeepSeek V4 Flash and the model it replaces, Qwen 3.6-27B, which drops to #6.

What changed between the generations is not the parameter count — 27B to 27.8B — but what the model does with it. Qwen3.8-27B is a native vision-language model that takes images and video, runs a 262,144-token context natively (1M with YaRN), and exposes a thinking mode with three reasoning_effort levels. The model card reports SWE-Bench Pro 61.7 and Terminal-Bench 2.1 73.0 — both on the vendor’s own scaffold, which the index annotates because vendor scaffolds typically run 15–20 points above standardized harnesses — plus LiveCodeBench v6 90.3 and GPQA Diamond 89.2.

For a Mac buyer the practical facts are unchanged from the 3.6 generation: about 19 GB at 4-bit, so it fits a 24 GB machine, and roughly 29 tok/s estimated on an M5 Max. The ecosystem caught up fast — over a thousand GGUF variants are listed, and Ollama, llama.cpp and LM Studio all run it.

GLM-5.3-Flash — the Ox Alpha Model, MIT, on a Mac Studio

For most of August an unattributed model called “Ox Alpha” was drawing attention on evaluation arenas. On 26 August Zhipu identified it as GLM-5.3-Flash and released the weights under MIT on Hugging Face.

The shape is the story. 320B total parameters, 18B active per token, the first natively multimodal model in the GLM-5 series, and a hybrid of sparse and linear attention that the card says cuts long-context serving cost. The one benchmark on the card is Terminal-Bench 2.1 at 84.3 — for scale, Qwen3.8-27B reports 73.0 on the same benchmark. Zhipu’s own line is that it outperforms GLM-5.2 across its benchmarks and approaches Claude Opus 4.8 on coding and agentic work. No SWE-Bench Pro figure was published on the card as of 7 September, so the index scores it at 48 on the evidence it has.

Whether it fits a Mac comes down to quantization, and Unsloth’s dynamic GGUFs make the tiers explicit:

Two caveats. The GGUFs need a GLM-5-aware llama.cpp branch as of this writing rather than a stock release, and the MLX route is community conversions. And the speeds are estimates using the index’s new MoE factor; a 3-bit quant on a 128 GB machine leaves 8 GB for everything else, which is a real constraint on context length. It ranks #10 with a Score of 68 because the 128 GB floor costs it accessibility points — by design. Capability alone would put it near the top.

And the flagship: GLM-5.3

Flash was not Zhipu’s only release. GLM-5.3 — the post-training upgrade of GLM 5.2 on the same 753B mixture-of-experts base, about 40B active — went out through Zhipu’s API on 18 August, and the weights followed on 25 August, the day the index last cut its data, which is how it was missed until this edition. The card reports Terminal-Bench 2.1 at 88.2 and DeepSWE v1.1 at 66.9; the headline gain over 5.2 is Terminal-Bench 3.0 moving from 4.6 to 28.3.

The license is the part to read. GLM 5.2 is MIT. GLM-5.3 ships under a bespoke GLM-5.3 License: commercial use permitted, attribution required, and one gate — model-as-a-service operators whose aggregate revenue exceeds US$10 billion over twelve months must pass Z.AI’s security review. For a Mac owner, or nearly any business, that is MIT in practice, and the index scores it 9/10, one point under MIT, in the same tier as Kimi’s Modified MIT. It is the first time Zhipu has split its licensing between a Flash and a flagship, and the direction is worth noting. A 2-bit quant is 239 GB, so it starts at a 256 GB Studio — about 29 tok/s estimated on an M5 Ultra — and it enters the index at #8 with a capability score of 50.

Qwen3.8-Flash-Next — 6B Active and a License Clause to Read

Released the same day as GLM-5.3-Flash, Qwen3.8-Flash-Next is the one Alibaba describes as a preview of the architecture that will underpin Qwen4: 125B parameters with 6B activated per token, plus a 51B n-gram embedding and a 4B multi-token-prediction head; Gated DeltaNet paired with Qwen Sparse Attention across 48 layers; 512 experts of which 10 routed and 1 shared fire per token. Native text, image and video input. A 262K context, 1M with YaRN. Thinking on by default with the same three effort levels as the 27B.

The card reports SWE-Bench Pro 62.5 (vendor scaffold), LiveCodeBench v6 91.9 and GPQA Diamond 91.7 — each a little above Qwen3.8-27B — and the index scores it at 47.

The number that matters for a Mac is 112 GB: the size of the community 4-bit MLX conversion. The 51B n-gram table lives in memory with everything else, so this is a 128 GB-tier model, not a 96 GB one. At that tier it is fast for what it is — about 49 tok/s estimated on an M5 Max, 78 on an M4 Ultra, 85 on an M5 Ultra — because only 6B parameters stream per token. The generators that build our per-Mac pages now pick it as the top model on every 128 GB-and-up configuration, from the MacBook Pro M5 Max to the 512 GB Studio. It ranks #13 overall because, like GLM-5.3-Flash, the RAM floor costs it points.

DeepSeek V4 Flash Vision-Exp

DeepSeek ran V4-Flash-Vision-Exp API-only from 21 to 30 August, then published the full weights on 31 August under plain MIT. It is V4 Flash with visual modules added and continued training on top — 305B total against the text model’s 284B, the same 13B-active decode path — and the card calls it experimental in so many words.

It reaches a Mac through Unsloth GGUFs: 2-bit at 91–97 GB for a 128 GB machine, 4-bit at 137–155 GB for a 192 GB Studio. Image input needs llama.cpp b10766 or later. The index lists it at #3 with the text model’s capability score of 47, because no independent vision benchmark has been verified yet, and estimates its speed by anchoring to V4 Flash’s community-measured 39 tok/s on M5 Max — same decode path, same number, labelled estimated until someone measures the vision checkpoint itself.

Muse Spark 1.3 — Still Closed

Meta released Muse Spark 1.3 on 2 September into its API and the Muse Code CLI, with the pitch that it asks clarifying questions, confirms before consequential actions, and cuts tool calls by about a fifth and tokens by about a quarter. It is not open weights. Mark Zuckerberg said open weights are coming “soon,” but what Meta has actually committed to releasing is version 1.2, without a date or license — and reporting this month put 1.2 behind Kimi K3, GLM-5.3 and Qwen3.8 on aggregate benchmarks. The index does not list a model until the weights exist. Muse Glimmer 30B, Meta’s Apache 2.0 release from July, remains the only open Muse.

For context, the closed frontier moved the same week: GPT-6 Astra on 3 September, Gemini 3.8 Flash on 2 September, Claude Fable 5.1 on 1 September. None of them affects a Mac leaderboard, except as the target the open models are chasing.

Calibration Note — Why a Published Figure Changed

The index estimates speed with a published two-term memory-bandwidth model fitted to vendor-measured dense runs. Sparse models need one more step. Their decode reads only the active parameters, which the naive formula rewards generously — it would put a 6B-active model at roughly 120 tok/s on an M5 Max. The index’s one community-measured MoE row, DeepSeek V4 Flash at 39 tok/s on M5 Max, comes in at 0.42 of that formula, and the MoE estimates already in the index all sat between 0.24 and 0.53 of it. Routing, expert loading and KV-cache traffic do not scale down with active size.

So from this edition every MoE estimate is the active-bytes formula scaled by 0.42, and that factor is documented on /methodology. Applying it consistently exposed one existing row that had been set with the unscaled formula: DeepSeek V4 Flash on M5 Ultra, published at ~80 tok/s in our 25 August coverage, is now ~47. The two Mac Studio posts that quoted the old figure carry a dated note. If you bought a 256 GB Studio on the strength of 80, the honest number is 47 until a measured run says otherwise, and a measured run is exactly what we want — the machine ships on 22 September.

Open-Source Top 10 (September 2026)

Ranked by LLMCheck Score, which weights capability alongside real Mac runnability — speed, RAM tier, license. The full 85-model table is on the leaderboard.

By Mac Tier — What to Run Today (September 2026)

Per-Mac rankings for all 44 configurations are on Best LLM by Mac; the Mac Advisor turns a budget into a config.

What’s Coming Next Month

Speculative section. Each item carries its status; nothing here enters the index until weights exist on a primary source.

• Announced Muse Spark 1.2 weights. Promised again on 2 September, still no date or license. Version 1.3 is not part of the commitment.
• Rumored Qwen4. Alibaba calls Flash-Next a preview of the Qwen4 architecture, which is the closest thing to a roadmap it has published. No name, size or date.
• Released Mac mini M6, Mac Studio M5 Max and M5 Ultra ship on 22 September. Every M6 and M5 Ultra figure in the index is an estimate today; the first independent runs will replace them, and the M5 Ultra rows for the three models above are the ones we most want measured.
• Rumored Nemotron 4 and Mistral’s frontier MoE. Both on August’s watch-list; nothing public in the window.

Frequently Asked Questions