GLM-5.3-Flash on a Mac: Which Studio Runs the Ox Alpha Model (2026)

What GLM-5.3-Flash is

On 26 August 2026 Zhipu confirmed that the model circulating as “Ox Alpha” was GLM-5.3-Flash and published the weights on Hugging Face under MIT — the most permissive license in the index, scored 10/10.

From the model card: 320B total parameters, 18B active per token, a hybrid of sparse and linear attention with what Zhipu calls manifold-constrained hyper-connections, native image input alongside text, and a 300K-token evaluation context. It is the first natively multimodal model in the GLM-5 series. The card reports Terminal-Bench 2.1 at 84.3; Zhipu’s launch line is that it outperforms GLM-5.2 across benchmarks at a tenth of the price and approaches Claude Opus 4.8 on coding and agentic tasks. No SWE-Bench Pro number appears on the card as of 7 September, so the index scores capability at 48 on the evidence available — one point under GLM-5.2, which it has not yet been shown to beat on a standardized harness.

Its sibling shipped alongside it: the 753B GLM-5.3 flagship, also open-weight since 25 August but under a custom license rather than MIT, and a 256 GB Studio proposition at 2-bit. This post is about the one you can run on 128 GB.

Which Mac holds which quant

Unsloth’s dynamic GGUFs are the practical reference. The sizes below are theirs; the Mac tiers are the index’s.

The line that matters: 128 GB is the floor, and 256 GB is where you stop compromising. A 96 GB Mac cannot hold any of these.

Speed per chip

The figures below come from the index’s published bandwidth model and are refilled from the index on every build, with a mark saying what each one is. For mixture-of-experts models above 40B total parameters, the model scales active-parameter bytes by a factor of 0.42. Treat that as a rough guide rather than a floor: measured runs of large MoEs have landed both above and below it.

The index costs the 128 GB row at 3-bit and the Ultras at 4-bit, which is why a 192 GB M2 Ultra can show less than a 128 GB M5 Max: it reads more bytes per token.

The first independent figure. One user’s run on a 256 GB M5 Ultra, summarised by llamaperf on 11 October, measured 39.7 tok/s plain decode in llama.cpp with a 262K context allocated, about 15% above the ~35 estimate for 4-bit. Its quantization is not stated and it may have used a branch build, so the index lists it as a community row but does not scale other chips from it. With GLM’s three-token MTP head switched on, the same run reached about 61 tok/s; multi-token prediction is a different workload, so that figure stays out of the dataset.

On a 128 GB M5 Max there is little room left at 3-bit. A community oMLX run of a 2.5-bit build with a third of its experts streamed from SSD measured about 13 tok/s at a 1K context. Streaming from SSD is a different workload from a model held in memory, so that figure is cited here and not listed as a row.

For comparison, DeepSeek V4 Flash, 13B active and the closest sibling in the index, is community-measured at 39 tok/s on an M5 Max at 2-bit. Rows live on the benchmarks page with their labels.

How to run it today

Since 30 September stock llama.cpp runs GLM-5.3-Flash; the branch build this guide first described is no longer needed.

llama.cpp with the GGUFs

Support merged in PR #27773 (build b11279) and shipped in llama.cpp 0.6.0 on 5 October; multi-token prediction followed in PR #29928 (build b11474), which shipped in 0.7.0 on 15 October. Homebrew’s llama.cpp is at 0.7.0, so on a Mac the shortest path is:

brew install llama.cpp
llama-cli -hf unsloth/GLM-5.3-Flash-GGUF:UD-IQ3_XXS -c 16384

Use files converted for the merged glm5-next architecture. Files made for the earlier support branch named it glm5next and will not load; Unsloth rewrote the first, 9 MB shard of every quant on 5 October to fix that, so a copy downloaded before then needs only that shard replaced. On a 256 GB Studio, swap UD-IQ3_XXS for UD-Q4_K_XL. Unsloth’s own guide still points at its branch; stock builds from b11279 on read the same files.

MLX and oMLX

Stock mlx-lm does not support the glm5_next architecture. Two MLX routes do: oMLX, whose maintainer publishes oQ builds from about 2-bit (110 GB, the one that fits a 128 GB Mac) to 6-bit, and mlx-vlm, which the mlx-community 4-bit build (204 GB, for a 256 GB Mac) targets. Check that a build’s size matches the Mac tier above before downloading 200 GB.

Ollama

Ollama’s official glm-5.3-flash entry is cloud-only (glm-5.3-flash:cloud): it runs on Ollama’s servers and does not download the weights. For a local run, use llama.cpp or oMLX.

Should you buy a Mac for it?

Not on this model alone — but this model is the clearest case yet that the 256 GB Mac Studio tier has a job. If you already own a 128 GB machine, the 3-bit quant is worth an evening. If you are deciding between 128 and 256 GB, GLM-5.3-Flash at true 4-bit, Qwen3.8-Flash-Next with room to spare, and DeepSeek V4 Flash at 4-bit are the three models that only the larger tier runs properly, and all three arrived in the same fortnight. The M5 Ultra Studio shipped on 22 September; the index’s per-Mac rankings are on the 256 GB page.

Frequently Asked Questions