MacLocalModels

Agent benchmark × Mac fit

Which Terminal-Bench open models have Mac-sized artifacts?

A narrow evidence join: official agent scores, canonical Hugging Face weights, exact GGUF footprints, and unverified arithmetic estimates of which Apple Silicon memory tiers leave room beyond the files.

Current answer

Two exact open-weight matches. Both exceed 216GB at the smallest observed quant.

GLM-5.3 and GLM-5.2 are the only exact Hugging Face models displayed on the official Terminal-Bench 3.0 or 4.0 leaderboards at capture time. Their Mac tiers below are unverified file-fit candidates—not launch or performance claims.

Download crawlable JSON ↗

Five-score reliability curve

One run answers capability. Five runs expose retry value.

For Terminal-Bench 4.0, Pass@1 through Pass@5 estimate whether at least one attempt succeeds. They are not five independent benchmarks and should not be compared with Mac token speed.

Artificial Analysis references

Inspect hosted intelligence, speed, and economics at source.

Read methodology ↗

Artificial Analysis maintains exact comparison pages for both models. Its current metric values are not republished here. Use those pages for hosted-provider context, then return to the Hugging Face and Mac artifact evidence below. Hosted performance is not Mac-local performance.

GLM-5.3Open exact model analysis ↗composite intelligence · hosted output speed · hosted API pricing · context window · model sizeGLM-5.2Open exact model analysis ↗composite intelligence · hosted output speed · hosted API pricing · context window · model size

Z.ai · exact open-weight match

GLM-5.3

TB 4.0 rank#9
Pass@141.8%
Pass@250.8%
Pass@354.7%
Pass@456.7%
Pass@557.6%

Five benchmark checkpoints

Terminal-Bench 4.041.8%Official leaderboardTerminal-Bench 3.032.4%Official leaderboardTerminal-Bench 2.188.2%Upstream model cardDeepSWE 1.166.9%Upstream model cardSWE-Marathon 1.142.5%Upstream model card

Unverified artifact-fit estimate

512GB fit candidate

Smallest observed quant
216.7GB · UD-IQ1_S (6 shards)
Lower-loss observed quant
281.7GB · UD-IQ3_XXS (7 shards)
256GB arithmetic
256GB Mac Studio; 39.3GB remains before runtime, system, and KV cache
512GB arithmetic
512GB Mac Studio; 230.3GB remains before runtime, system, and KV cache
Calculation
Decimal unified-memory tier minus total decimal GGUF shard bytes; no runtime overhead or KV-cache reservation is included.
Runtime candidate
llama.cpp-compatible GGUF runner; current GLM-5.3 support must be confirmed

Not launched or measured on a Mac. Artifact fit does not prove runtime support, speed, context capacity, or benchmark retention.

Z.ai · exact open-weight match

GLM-5.2

TB 4.0 rankNot listed
Pass@1—
Pass@2—
Pass@3—
Pass@4—
Pass@5—

Five benchmark checkpoints

Terminal-Bench 4.0Not listedNo official entryTerminal-Bench 3.04.6%Official leaderboardTerminal-Bench 2.181%Upstream model cardDeepSWE upstream model card46.2%Upstream model cardSWE-Marathon upstream model card13%Upstream model card

Unverified artifact-fit estimate

512GB fit candidate

Smallest observed quant
216.7GB · UD-IQ1_S (6 shards)
Lower-loss observed quant
281.7GB · UD-IQ3_XXS (7 shards)
256GB arithmetic
256GB Mac Studio; 39.3GB remains before runtime, system, and KV cache
512GB arithmetic
512GB Mac Studio; 230.3GB remains before runtime, system, and KV cache
Calculation
Decimal unified-memory tier minus total decimal GGUF shard bytes; no runtime overhead or KV-cache reservation is included.
Runtime candidate
llama.cpp-compatible GGUF runner; current GLM-5.2 support must be confirmed

Not launched or measured on a Mac. Artifact fit does not prove runtime support, speed, context capacity, or benchmark retention.

How to use this data

Choose capability first, then test the exact local artifact.

  1. Use Terminal-Bench to assess long-horizon terminal-task evidence for the named model-agent configuration.
  2. Use the Hugging Face links to identify the exact weights and all shards required by one quantization.
  3. Leave memory for the runtime and KV cache; file size alone is not enough.
  4. Rerun representative tasks locally before relying on a quantized model for agent work.

Captured 2026-09-14. Terminal-Bench versions contain different tasks and are not directly comparable. Model-card scores are labelled separately from official leaderboard results. Artificial Analysis hosted-provider observations are kept separate from local Mac evidence.