Qwen-AgentWorld-35B-A3B-heretic

A capability-preserving, uncensored ("abliterated") version of Qwen/Qwen-AgentWorld-35B-A3B, produced with Heretic.

Base model: Qwen/Qwen-AgentWorld-35B-A3B โ€” 35B parameters (3B active). This checkpoint grafts the Qwen3 VL vision tower and Multi-Token Prediction (MTP) tensors back on, so it ships vision weights and MTP heads despite the text-only base. Mixture-of-Experts (40 layers, 256 routed experts, ~3B active per token), 262,144-token context.

Weights only; architecture, tokenizer, and chat template are unchanged.

Uncensored model. This checkpoint complies with a wider range of requests than the base model. You are responsible for your use of it.

Benchmarks (AgentWorldBench, Token F1)

BF16 baseline:

Domain Score
Terminal 50.5%
Search 12.1%
SWE 51.8%
MCP 77.5%
Android 32.0%
Overall 44.8%

Format compliance: 100% across all domains with <predicted_observation>...</predicted_observation> prompting. EXL3 quants of this checkpoint retain 97-100% of the BF16 AgentWorld F1.

License & attribution

Released under Apache-2.0, inheriting the base model's license. Abliteration performed with Heretic. Base model and all underlying credit: the Qwen team.

Downloads last month
100
Safetensors
Model size
35B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for genevera/Qwen-AgentWorld-35B-A3B-heretic

Finetuned
(11)
this model

Collection including genevera/Qwen-AgentWorld-35B-A3B-heretic