A 744B Open Model Beats Opus on AIME, If You Own $50k of GPUs

Z.ai's GLM-5.2 is a 744B-total, 40B-active MoE built for long-horizon coding and agentic work, with a 1M-token context. Unsloth's guide frames it as a frontier-class open model you can run at home, and the benchmarks back the first claim: it edges Opus on competition math while trailing on real-world software tasks.

How GLM-5.2 stacks up against the proprietary frontier

| Benchmark | GLM-5.2 | Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro | |---|---|---|---| | AIME 2026 | 99.2 | 95.7 | 98.3 | 98.2 | | SWE-bench Pro | 62.1 | 69.2 | 58.6 | 54.2 |

The pitch hinges on Unsloth's Dynamic GGUF quantization, which it says keeps a 1-bit build only ~24% less accurate than the full 1.51TB model. The memory ladder:

The only at-home benchmark in the thread came from someone running it for real.

"I run Q4_K_XL. All it takes … is 512gb of ram and 2 3090 GPUs with llama.cpp -cmoe… about 6tk/sec… I did a budget build before the crazy hardware cost and I regret it everyday. Nevertheless, it's fantastic being able to run this model at home." — a builder running it on a 512GB DDR4 + dual-3090 rig

The slow path has a hidden upside. Bitcoin Core's Greg Maxwell found that CPU-only inference scales sideways:

"Just running cpu only w/ Q6 on 9684X I get about 1tok/s … also still get about 1tok/s/stream when running 16 in parallel." — Greg Maxwell, Bitcoin Core developer

But the most pointed correction targets the word "local" itself.

"'it can fit' on 256GB of RAM, but it will be heavily quantized and still run very slowly. The headline number is not token generation, its prompt processing… a mac studio… will do PP 20-50X slower… which is what actually makes this unusable without $50k in GPUs." — a commenter separating prompt-processing from token-generation

Even the accuracy framing drew an audit, after one reader checked Unsloth's own chart:

"It says 'dynamic 4-bit … generally lossless' while showing a top-1% token agreement on the chart of 97.5%. Not what I would consider 'generally lossless'." — a commenter auditing the source's chart

The model genuinely runs on prosumer hardware; what the guide elides is that "runs" means tolerable generation speed long after prompt processing has made the machine unusable for the agentic workloads GLM-5.2 was built for.