GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
darrellbest
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
darrellbest
Model Tree
darrellbest/Qwen3.8-27B-Heretic
Input Modalities
Output Modalities
Supported Functionality
| Part | Precision |
|---|---|
| Linear layers of the MLPs and the 16 full-attention layers | NVFP4, 16-value groups, FP8 scales |
Gated DeltaNet (linear_attn) layers, vision tower, embeddings, lm_head, MTP | bf16 (unchanged) |
This model is 64 layers of which 48 are Gated DeltaNet, and those keep a recurrent state that low precision damages,
so they stay in bf16 along with the vision tower. That is why the file is 27 GB rather than roughly 15 GB.
Multi-token-prediction weights are preserved in model-auxiliary.safetensors, which save_pretrained drops for this
architecture.
Made with llm-compressor 0.13.0 (scheme="NVFP4"), calibrated on
64 harmless chat prompts. Format: compressed-tensors, nvfp4-pack-quantized.
sh
vllm serve darrellbest/Qwen3.8-27B-Heretic-NVFP40/100 refusals against 98/100 for the original, KL divergence 0.0465. Method and parameters are on the parent card. The 4-bit weights were not re-measured for refusals, and the 4-bit activation path is vLLM's, which was not exercised here.
Reduced safety guardrails by design. You are responsible for what you do with it.
| Repository | Format | Size | Use it with |
|---|---|---|---|
| Qwen3.8-27B-Heretic | bf16 safetensors | 51 GB | transformers, vLLM, SGLang |
| Qwen3.8-27B-Heretic-GGUF | GGUF BF16 / Q8_0 / Q4_K_M (+ vision) | 51 / 27 / 16 GB | llama.cpp, Ollama |
| Qwen3.8-27B-Heretic-FP8 | FP8 W8A8, compressed-tensors |
| 35 GB |
| vLLM |
| Qwen3.8-27B-Heretic-NVFP4 | NVFP4, compressed-tensors | 27 GB | vLLM on Blackwell |