GLM-5.2 · BF16 · full precision

GLM‑5.2 BF16

The full-precision reference weights — the refusal-ablated GLM‑5.2 exactly as delivered. Fine-tune it, distill from it, or roll your own quants for evaluation and detector work.

Red-team baseline · 0/450 bfloat16 · full precision 512K → 1M Fine-tune ready
One-time · commercial + research
$499
Buy GLM‑5.2 BF16Bitcoin — $449 · save 10%
Secure checkout via Stripe
Gated research access approved after purchase
Overview

The source of truth for your own research models.

GLM‑5.2 BF16 is the unquantized, full-precision checkpoint — the same refusal-ablated GLM‑5.2 base behind the NVFP4 builds, delivered raw. This is the build for teams that need to fine-tune a domain-specific red-team model, distill a smaller evaluation model, or produce their own quantizations (GGUF, AWQ, custom NVFP4). Full MoE, DSA sparse attention, native long context, a clean 0-refusal baseline to build on.

Specifications
BaseGLM-5.2 (refusal-ablated)
ArchitectureMoE · DSA sparse attn
Experts256 / 256 (full)
Precisionbfloat16 (unquantized)
Context512K (→ 1M, YaRN)
Refusal benchmark0 / 450 adversarial
Weights~1.5 TB
Formatsafetensors · HF Transformers
Best forfine-tune · distill · quantize
Licensecommercial + research
What's included
  • Full BF16 weights — gated Hugging Face research download
  • config, chat template, tokenizer — ready to load
  • Quantization notes — the NVFP4 + MTP recipe pointers
  • Commercial + research license, including for derivatives you train
Quickstart

Load it, tune it, evaluate.

# load the reference weights
from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained("glm-5.2-bf16", dtype="bfloat16", device_map="auto")
# ... then fine-tune a red-team model, distill an evaluator, or quantize to your target format
Who it's for

Fine-tuners

Adapt the base into a domain-specific red-team or evaluation model from a clean full-precision start.

Quant builders

Produce your own GGUF / AWQ / NVFP4 research targets from the source weights.

Researchers

Distillation, interpretability, and safety evaluation on the unquantized model.

Other builds

Just want to serve it?

The NVFP4 builds are ready to deploy — full on eight cards, or REAP-pruned on four.

GLM-5.2 REAP (4-card) $399 → GLM-5.2 NVFP4 (Full) $499 →