Downloads · 30 days
20
4% of all-time downloads
maltsevis/Qwen-AgentWorld-Heretic-HCl-NVFP4-dgx
Qwen-AgentWorld-Heretic-HCl-NVFP4-dgx is a text generation model from maltsevis. Use it when you need the model to write or continue text. The card lists the license as apache-2.0.
<div align="center" <img src="https://huggingface.co/datasets/Qwen/AgentWorldBench/resolve/main/assets/logo.png" width="400px" <h1Qwen-AgentWorld-Heretic-HCl-NVFP4</h1 <p<em35B MoE · NVFP4 4-bit · Heretic-abliterated…
Downloads · 30 days
20
4% of all-time downloads
All-time downloads
479
Public
Parameters
20B
21 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors21 GB · 100%
How the weights are stored.
U816.8B · 84%
From the Hugging Face model README
| Base model | Qwen/Qwen-AgentWorld-35B-A3B |
| Architecture | Qwen3.5 MoE — 35B total / 3B active |
| Experts | 256 experts, top-8 routing |
| Layers | 40 transformer (linear attention + full attention) |
| Context | 262,144 tokens (256K) |
| Quantization | — 4-bit NVIDIA floating point (e2m1) |
| Compression | 21 GB (3.1× from 65 GB bf16) |
| Format | safetensors, vLLM-compatible |
This model underwent controlled steering via Heretic v1.4.0 to reduce refusal behavior while minimizing KL divergence.
| Parameter | Value |
|---|---|
| Trials | 200 (completed) |
| KL divergence target | |
| Good prompts | (400) |
| Bad prompts | (400) |
| Row normalization | FULL, LoRA rank 3 |
| Refusal markers | 28 |
| GPU | NVIDIA GB10, 121 GB |
Custom pipeline on DGX Spark (GB10, 121 GB unified memory):
| Step | Detail |
|---|---|
| Load | (half the memory of bf16) |
| Calibration | 32 samples, |
| Quantization | 0.44, |
| Export | |
| Remap | for vLLM |
| Component | Meaning |
|---|---|
| Qwen | Base architecture family |
| AgentWorld | Language world model — 7 simulation domains |
| Heretic | Abliterated via Heretic parameter study |
| HCl | Production mode (hydrochloric acid) |
| NVFP4 | 4-bit NVIDIA floating point quantization |
| dgx | Quantized & deployed on DGX Spark |