How to Run GLM 5.2 Air on a Mac: Frontier-Class AI on Apple Silicon (July 2026)

What GLM 5.2 Air Actually Is

GLM 5.2 Air is the compact sibling of GLM 5.2, Zhipu AI's flagship reasoning model. The full GLM 5.2 is a 744B-parameter mixture-of-experts (MoE) model with 40B active parameters per token — a model that lives on multi-GPU server racks, not laptops. Air is what you get when that flagship is distilled down to a size that runs on consumer Apple Silicon.

The Air configuration is 106B total parameters with only 12B active at any given step. That MoE design is the whole trick: the model holds 106B parameters' worth of knowledge in memory, but each generated token only routes through 12B of them. You pay the RAM cost of a large model while paying the compute cost of a small one — which is precisely what makes it viable on a Mac.

It ships with a 256K-token context window and the MIT license, the most permissive option in the open-weights world. According to LLMCheck benchmarks, it lands at 58% on SWE-Bench Pro, 88% on MMLU, and 92% on HumanEval — numbers that would have been the domain of closed frontier APIs a year ago.

Hardware Requirements

At the default Q4_K_M quantization, GLM 5.2 Air's weights occupy roughly 56 GB of unified memory. Add the OS footprint, your apps, and a working context buffer, and you need a machine with at least 64 GB. Here is how the tiers shake out:

A practical note for 64 GB owners: the model fits, but unified memory is shared with everything else on the system. Quit memory-hungry apps before loading the model, and cap your context length to avoid swapping. The M4 Ultra and M5 Max with 128 GB+ are where GLM 5.2 Air feels effortless.

Speeds by Apple Silicon Chip

Because only 12B parameters are active per token, GLM 5.2 Air punches far above what a dense 106B model could ever manage on a Mac. According to LLMCheck benchmarks, measured throughput across chips looks like this:

All four configurations sit comfortably above conversational reading speed (~10–15 tok/s). The M4 Ultra's wider memory bus and the M5 Max's improved per-core throughput push it past 34 tok/s, where responses feel near-instant. Even the 64 GB M5 Max — the cheapest viable machine — holds a steady 30 tok/s, which is more than enough for interactive coding and chat.

Step-by-Step Install (Ollama + MLX)

There are two clean paths to running GLM 5.2 Air locally. Ollama is the fastest to set up; MLX squeezes out a few extra tokens per second on Apple Silicon.

Option A — Ollama (easiest)

If you do not already have Ollama, install it from the software page or via Homebrew, then pull and run the model in a single command:

That single ollama run command downloads the Q4_K_M weights, loads them into unified memory, and drops you into an interactive prompt. To serve it to other apps over the local API, run ollama serve and point your client at http://localhost:11434.

Option B — MLX (fastest on Apple Silicon)

MLX is Apple's own array framework, tuned for the unified-memory architecture. It typically delivers a few extra tok/s at the same quantization. Install mlx-lm and pull the community-quantized build:

For an interactive chat loop or an OpenAI-compatible server, use mlx_lm.chat or mlx_lm.server respectively. The first invocation downloads the weights from Hugging Face; subsequent runs load from the local cache.

Recommended Quantization

Quantization is the lever that trades memory and speed against output fidelity. For GLM 5.2 Air on a Mac, two settings cover almost every case:

• Q4_K_M (default) — ~56 GB of weights. The right choice on a 64 GB Mac and the standard build shipped by both Ollama and the MLX community. Quality loss versus full precision is small and rarely noticeable in coding or chat.
• Q5_K_M — ~68 GB of weights. Worth it only on 128 GB+ machines, where the extra headroom is free. It tightens up reasoning on the hardest multi-step problems and reduces occasional formatting drift in long agentic runs.

Skip Q8 and full precision on a single Mac — the memory cost is not justified by the marginal quality gain for this model. If you are on exactly 64 GB, stay at Q4_K_M and keep your context window in check rather than chasing a higher quant.

Real Workflows: Agentic Coding & 256K RAG

GLM 5.2 Air's combination of strong code scores and a 256K context window unlocks two workflows that smaller local models struggle with.

Agentic coding

With 92% on HumanEval and 58% on SWE-Bench Pro, Air is genuinely useful as a local coding agent. Wire it into an editor or an agent framework via Ollama's OpenAI-compatible endpoint, and it can plan a change, edit multiple files, and reason about test failures. At 30 tok/s the intermediate steps of an agent loop accumulate quickly, but the model rarely loses the thread — the distillation preserves the flagship's tendency to think before it acts.

Whole-repo RAG over 256K context

The 256K-token window is the headline feature for retrieval-augmented work. You can load an entire mid-sized codebase, a long technical spec, or a stack of documentation directly into context and ask the model to reason across all of it at once — no chunking, no embedding pipeline required for many tasks. On a 128 GB machine you have the memory to actually use that full window; on 64 GB, keep the loaded context to a few tens of thousands of tokens and let a retrieval step pull in only the relevant slices.

GLM 5.2 Air vs Qwen 4.1 32B-A3B

The natural alternative on a Mac is Qwen 4.1 32B-A3B — a smaller MoE that runs on far less RAM. The two solve different problems:

Limitations to Know

GLM 5.2 Air is impressive, but it is a distillation — and distillations lose something. Set expectations accordingly:

• It is not the full GLM 5.2. On SWE-Bench Pro it scores 58% versus the flagship's 68.5%. On the hardest agentic and multi-step reasoning tasks, that ~10-point gap is real — Air occasionally takes a wrong turn where the 744B model would not.
• 64 GB is genuinely the floor. The model fits, but it leaves little room for large contexts or other heavy apps. Memory pressure and swapping will tank your throughput if you are not disciplined.
• The 256K context is RAM-bound. Filling the full window costs unified memory on top of the weights. Only 128 GB+ machines can use it without compromise.
• MoE routing adds variance. Because different tokens activate different experts, throughput can fluctuate slightly run to run more than a dense model would.

None of these are dealbreakers. For a model that brings frontier-distilled reasoning to a single Mac under the MIT license, GLM 5.2 Air is the most capable local option in its class as of July 2026.