---

base_model: Qwen/Qwen-AgentWorld-35B-A3B

base_model_relation: quantized

quantized_by: ivgranite

pipeline_tag: text-generation

license: apache-2.0

license_link: https://huggingface.co/Qwen/Qwen-AgentWorld-35B-A3B/blob/main/LICENSE

tags:

• gguf
• qwen3_5_moe
• world-model
• agent
• environment-simulation
• base_model:Qwen/Qwen-AgentWorld-35B-A3B
• base_model:quantized:Qwen/Qwen-AgentWorld-35B-A3B
• conversational
• endpoints_compatible

language:

• en

library_name: gguf

---

Qwen-AgentWorld-35B-A3B GGUF

GGUF quantization of Qwen/Qwen-AgentWorld-35B-A3B, a native language world model trained for agentic environment simulation across seven domains (Terminal, MCP, Search, SWE, Web, Android, OS).

Why this conversion exists

Most community GGUF conversions of this model have a critical bug: they set block_count=41 (inheriting mtp_num_hidden_layers: 1 from the Qwen3.5 config template), but the original safetensors contain zero MTP/nextn tensors. This causes llama.cpp to fail with:

error loading model: missing tensor 'blk.40.attn_norm.weight'

This conversion was done from the original safetensors using llama.cpp's convert_hf_to_gguf.py with --no-mtp, which correctly sets block_count=40 and skips the non-existent MTP layer.

For additional quants (Q2_K, Q6_K, Q8_0, BF16, imatrix), groxaxo's conversion is also correct and worth checking out.

Conversion details

Toolchain: llama.cpp release b9484 (commit 63e66fdd2)

# Download original model
hf download Qwen/Qwen-AgentWorld-35B-A3B --local-dir ./source

# Convert to BF16 GGUF (no MTP — model doesn't include MTP weights)
python convert_hf_to_gguf.py ./source \
  --outfile agentworld-bf16.gguf \
  --outtype bf16 \
  --no-mtp

# Quantize
llama-quantize agentworld-bf16.gguf Qwen-AgentWorld-35B-A3B-Q4_K_M.gguf Q4_K_M

Verified: 733 tensors, block_count=40, no nextn_predict metadata.

Available quants

| Filename | Quant type | File Size | Description |

| -------- | ---------- | --------- | ----------- |

| Qwen-AgentWorld-35B-A3B-Q4_K_M.gguf | Q4_K_M | 19.7GB | Good quality, default for most use cases, recommended. |

Model info

• Architecture: Qwen3.5 MoE (Gated DeltaNet + Gated Attention hybrid)
• Parameters: 35B total / 3B active
• Layers: 40 (10 × [3 linear attention + 1 full attention])
• Experts: 256 (8 routed + 1 shared)
• Context: 262,144 tokens
• Thinking mode: Enabled by default (<think>...</think>)

Usage

# llama.cpp
llama-server \
  -m Qwen-AgentWorld-35B-A3B-Q4_K_M.gguf \
  --reasoning on --reasoning-budget 32768 \
  -c 131072 -ngl 999 -fa on \
  --cache-type-k q4_0 --cache-type-v q4_0 \
  --jinja

# Recommended sampling
temperature: 0.6, top_p: 0.95, top_k: 20

Prompt format

<|im_start|>system
{world_model_system_prompt}<|im_end|>
<|im_start|>user
{agent_action}<|im_end|>
<|im_start|>assistant
<think>

Domain-specific system prompts for all 7 domains (Terminal, MCP, Search, SWE, Web, Android, OS) are available in the official GitHub repository.

Evaluation

Run AgentWorldBench (2,170 samples, 7 domains) with the official eval framework:

git clone https://github.com/QwenLM/Qwen-AgentWorld.git
hf download Qwen/AgentWorldBench --repo-type dataset --local-dir ./AgentWorldBench

cd Qwen-AgentWorld/eval
python eval.py infer \
  --data-dir ../AgentWorldBench \
  --model-base-url http://localhost:8080/v1 \
  --model-name Qwen-AgentWorld-35B-A3B \
  --output-dir ./results

Original model

See Qwen/Qwen-AgentWorld-35B-A3B, the blog post, and the technical report (arXiv:2606.24597) for full details.