Model datasheet · GLM
GLM-4.7
The full-size GLM behind the 30B Flash already listed here, and the strongest open coding scores on this site — SWE-bench Verified 73.8. It is also the most expensive model here to hold in memory: all 92 layers use full attention, so the KV cache costs 368KB per token, more than any other model in the catalogue.
- Vendor
- Z.ai
- Architecture
- MoE · 358B total / ~32B active
- Context
- 202,752 tokens
- License
- MIT
- Released
- 2025-12-22
- Vision
- No
§1 Characteristics by quantization band
data checked Aug 2026
- quant-degradation-community — editorial band derivation @ Q3_K_M · aggregated · 2026-08-31
Editorial derivation: two points down. At 3-bit a MoE this large loses expert-routing precision first, which surfaces as dropped tool arguments and broken multi-file edits before prose visibly suffers.
data checked Aug 2026
- quant-degradation-community — editorial band derivation @ Q3_K_M · aggregated · 2026-08-31
Editorial derivation: two points down. At 3-bit a MoE this large loses expert-routing precision first, which surfaces as dropped tool arguments and broken multi-file edits before prose visibly suffers.
data checked Aug 2026
- quant-degradation-community — editorial band derivation @ Q4_K_M · aggregated · 2026-08-31
Editorial derivation: one point down from the measured bf16 figure. 4-bit is the band a model this size is actually run at, and the loss is real but modest.
data checked Aug 2026
- quant-degradation-community — editorial band derivation @ Q4_K_M · aggregated · 2026-08-31
Editorial derivation: one point down from the measured bf16 figure. 4-bit is the band a model this size is actually run at, and the loss is real but modest.
data checked Aug 2026
- quant-degradation-community — editorial band derivation @ Q4_K_M · aggregated · 2026-08-31
Editorial derivation: one point down from the measured bf16 figure. 4-bit is the band a model this size is actually run at, and the loss is real but modest.
data checked Aug 2026
- quant-degradation-community — editorial band derivation @ Q6_K · aggregated · 2026-08-31
Editorial derivation: no drop from the measured bf16 figure. 6-8 bit is effectively lossless on a model this size — the differences that show up at 4-bit and below are not measurable here. Recorded explicitly rather than left to band fallback, which would otherwise borrow the 4-bit number and subtract a point.
data checked Aug 2026
- quant-degradation-community — editorial band derivation @ Q6_K · aggregated · 2026-08-31
Editorial derivation: no drop from the measured bf16 figure. 6-8 bit is effectively lossless on a model this size — the differences that show up at 4-bit and below are not measurable here. Recorded explicitly rather than left to band fallback, which would otherwise borrow the 4-bit number and subtract a point.
data checked Aug 2026
- quant-degradation-community — editorial band derivation @ Q6_K · aggregated · 2026-08-31
Editorial derivation: no drop from the measured bf16 figure. 6-8 bit is effectively lossless on a model this size — the differences that show up at 4-bit and below are not measurable here. Recorded explicitly rather than left to band fallback, which would otherwise borrow the 4-bit number and subtract a point.
data checked Aug 2026
- quant-degradation-community — editorial band derivation @ Q6_K · aggregated · 2026-08-31
Editorial derivation: no drop from the measured bf16 figure. 6-8 bit is effectively lossless on a model this size — the differences that show up at 4-bit and below are not measurable here. Recorded explicitly rather than left to band fallback, which would otherwise borrow the 4-bit number and subtract a point.
FP16 — full precision (fp16/bf16)
717GB weights · e.g. bf16 + KV: 2.9GB at 8K · 11.5GB at 32K · 46GB at 128Kdata checked Aug 2026
- vendor-model-card — SWE-bench Verified = 73.8 @ bf16 · vendor-reported · 2026-08-31
LiveCodeBench-v6 84.9. The highest SWE-bench figure in this catalogue; Z.ai position 4.7 explicitly as a coding model.
data checked Aug 2026
- vendor-model-card — MMLU-Pro = 84.3 @ bf16 · vendor-reported · 2026-08-31
No hallucination-rate benchmark published, so this rests on a knowledge benchmark alone.
data checked Aug 2026
- vendor-model-card — AIME 2025 = 95.7 @ bf16 · vendor-reported · 2026-08-31
GPQA-Diamond 85.7, MMLU-Pro 84.3, HLE with tools 42.8.
data checked Aug 2026
- vendor-model-card — Tau2-Bench = 87.4 @ bf16 · vendor-reported · 2026-08-31
Terminal Bench 2.0 41.0, BrowseComp with context management 67.5. The Tau2 figure is the best agentic number on the site.
§2 Known issues & what fixes them
The most expensive KV cache in this catalogue by a wide margin. All 92 layers use full attention with no sliding window and no linear-attention layers, giving 368KB per token — so a 32K context costs 12GB of cache before a single weight is loaded, and filling the 198K window would cost about 71GB. On a machine sized to just hold the weights, context is what you run out of first.
Architectural. Budget the cache explicitly: the RUNS AT line on this site prices it at the context you pick.
Evidence · 2 sources · community-consensus
- GLM-4.7 config.json: 92 layers, no sliding_window, 8 KV heads x 128 head_dim (vendor, 2026-08-31)
- Unsloth GLM-4.7 run guide: memory requirements (blog, 2026-08-31)
For almost everyone this model is only reachable at 2-3 bit — the 4-bit build is 205GB and the bf16 weights are 717GB. Unsloth's dynamic 2-bit lands around 135GB, which is the first size that fits a single large machine. That is a real option, but it is the bottom of the quant ladder, and expert routing in a 358B MoE is exactly what degrades there.
Move up to the 205GB 4-bit build if you have the memory; the gap between 2-bit and 4-bit is larger here than on a small dense model.
Evidence · 2 sources · community-consensus
- unsloth GLM-4.7-GGUF: Q2_K_XL 135GB, Q3_K_M 171GB, UD-Q4_K_XL 205GB, BF16 717GB (other, 2026-08-31)
- Unsloth GLM-4.7 run guide (blog, 2026-08-31)