LocalClaw/LLM/Qwen AgentWorld 35B-A3B
Open-weight MoE

Qwen AgentWorld 35B-A3B

Official Qwen language world model for simulating agent environments across terminal, web, OS, Android, search, SWE and tool-calling domains. Apache 2.0 with active GGUF and MLX quantizations.

32 GB power user 32 GB RAM Q4_K_M Agent environment simulation
Run with LocalClaw Compare all models
Parameters
35B (3B active, MoE)
Minimum RAM
32 GB
Model size
20 GB
Quantization
Q4_K_M

Can Qwen AgentWorld 35B-A3B run locally?

Qwen AgentWorld 35B-A3B belongs on 32 GB machines when you want stronger quality without jumping to server hardware.

Search for qwen-agentworld-35b-a3b in LM Studio or another GGUF-compatible runtime.

Model sourceunsloth/Qwen-AgentWorld-35B-A3B-GGUF
chatcodereasoningagentpower

Install path

01
Check RAM fitMinimum 32 GB RAM. Start with the Q4_K_M quant.
02
Load the modelSearch qwen-agentworld-35b-a3b in LM Studio.
03
Control locallyUse LocalClaw to manage models, agents, chat, channels and scheduled OpenClaw work.

Strengths

  • Official Qwen Apache 2.0 release rather than a community fine-tune
  • Specialized for terminal, web, OS, Android, search, SWE and tool-calling environment simulation
  • Sparse 35B-A3B shape keeps active compute much lower than dense 35B models
  • Large 262K context window for multi-step agent traces
  • Strong public AgentWorldBench results for simulated environment observations
  • GGUF and MLX quantizations are already available for local testing

Limitations

  • Specialized world-model behavior, not a normal general chat default
  • Long-context use can raise memory requirements far above the small active-parameter count
  • Best served on 32GB+ local machines or GPU workstations with recent runtimes
  • Community quantizations should be validated before relying on tool or terminal simulation outputs

Best use cases

  • Agent environment simulation
  • Terminal and OS workflow rehearsal
  • Tool-calling and MCP test scenarios
  • SWE agent benchmark experiments
  • Long multi-step action trace prediction
  • Research on local agent training loops

Capability profile

speed
5
quality
8
coding
8
reasoning
9

Technical notes

Developer
Alibaba Cloud (Qwen Team)
License
Apache 2.0
Context window
262,144 tokens
Architecture
Qwen3.5 MoE-based language world model with 35B total parameters, about 3B active parameters, Gated DeltaNet, gated attention, 256 experts and 8 routed experts plus one shared expert per token.

This model fits these next steps

Hardware fit is based on LocalClaw's RAM tier, model size and quantization metadata. Always leave memory headroom for your OS and runtime.

Comfortable headroomMac mini M4 Pro 48GBMobile workstationMacBook Pro M4 Max 36GBPower-user picks32GB RAM guide

Similar models to compare

Qwen 3.5 MoE (35B/3B active) 35B (3B active)Qwen 3 MoE (30B/3B active) 30B (3B active)Kimi Linear 48B-A3B Instruct 48B (3B active, MoE)OLMo 3 32B Think 32B

Where to go next