Qwen-AgentWorld-35B-A3B
Qwen-AgentWorld-35B-A3B is a 35-billion parameter language model fine-tuned specifically to simulate agent environments across seven interaction domains (tool calling, search, terminal, software engineering, Android, web, and OS). It predicts the next state of an environment given an agent's action and interaction history, using extended context windows (262K tokens). The model is trained from inception as a 'world model' rather than adapted post-hoc. It runs on Apache 2.0 license, is ungated, and compatible with standard inference frameworks (vLLM, SGLang, Transformers).
Key facts
Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.
| Field | Value |
|---|---|
| Developer | Qwen |
| Parameters | 34.7B |
| Context window | Unknown |
| License | apache-2.0 — OSI-approved |
| Modality / task | text-generation |
| Gated on HuggingFace | No |
| Downloads | 60.7k |
| Likes | 560 |
| Last updated | 2026-06-25 |
| Source | Qwen/Qwen-AgentWorld-35B-A3B |
What Qwen-AgentWorld-35B-A3B is
Qwen-AgentWorld-35B-A3B is a Mixture-of-Experts (MoE) causal language model with 35B total parameters and 3B activated. Architecture: 40 layers, 2048 hidden dimension, 256 experts (8 routed + 1 shared). Uses Gated DeltaNet (linear attention) and Gated Attention layers in alternating pattern. Trained via three-stage pipeline: Continual Pre-Training (CPT, environment knowledge injection), Supervised Fine-Tuning (SFT, next-state reasoning), and Reinforcement Learning (GSPO). Context length 262,144 tokens. Base model: Qwen3.5-35B-A3B-Base. Compatible with Hugging Face Transformers, vLLM, SGLang, and OpenAI-compatible APIs.
Run Qwen-AgentWorld-35B-A3B locally
Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.
from transformers import pipelinepipe = pipeline("text-generation", model="Qwen/Qwen-AgentWorld-35B-A3B")out = pipe("Explain retrieval-augmented generation in one sentence.", max_new_tokens=128)print(out[0]["generated_text"])Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.
How you'd run it
A typical self-hosted path — open weights, an inference server, your application.
DEV.co builds each layer — from GPU infrastructure to the application.
Best use cases
Running & fine-tuning it
Estimated: 70–140 GB VRAM in float16 for 35B parameters with full 262K context (rough calculation: 35B × 2 bytes + KV cache overhead). Practical inference typically requires H100 (80GB) or multi-GPU setup with tensor parallelism (tp-size 4 shown in examples). For smaller context (128K minimum per docs), estimate ~50–80 GB VRAM. Quantization (int8, int4) may reduce to 40–60 GB but may impact world-model accuracy. Exact overhead depends on batch size and inference framework (vLLM, SGLang optimize differently).
Fine-tuning details for this specific model are not provided in the card. Base model is Qwen3.5-35B-A3B-Base (MoE architecture). Standard LoRA/QLoRA feasibility is unclear; MoE models may require adapter strategies that target expert or gate components. Recommend consulting Qwen documentation or community for MoE-specific fine-tuning. For agentic task adaptation, SFT on domain-specific trajectories is stated as part of the training pipeline; domain-specific system prompts provided in GitHub repo may guide supervised fine-tuning approaches.
When to avoid it — and what to weigh
- Single-GPU Inference at Full Context Length — The model's 262K context window and 35B parameters (3B active per step) require significant VRAM. Full-context inference on consumer-grade hardware is not practical; quantization or context truncation may degrade world-model fidelity per model card guidance.
- General-Purpose Conversational Chat — Qwen-AgentWorld is specialized for environment simulation and next-state prediction, not general-purpose conversation. It will underperform generic chat LLMs for non-agentic tasks. See benchmark comparison: GPT-5.4 scores 58.25 overall vs. Qwen-AgentWorld 56.39, and gaps widen outside agent domains.
- Real-Time Interactive Environments Requiring Sub-Second Latency — Extended context windows and MoE routing add latency compared to base models. For use cases requiring immediate environment response, optimization (token budget reduction, KV cache tuning) is necessary.
- Vision-Based Environment Simulation (Without Additional Adaptation) — Model card notes visual component definitions exist but checkpoint contains language-model weights only. Direct multimodal environment simulation (e.g., camera-based robot control) requires additional development or fine-tuning.
License & commercial use
Apache License 2.0. This is a permissive OSI-approved open-source license allowing commercial use, modification, and distribution under the license terms.
Apache 2.0 permits commercial use, including in production systems and proprietary products, provided you include a copy of the license and state material changes. No additional commercial license, proprietary restrictions, or per-seat fees stated in the card. The model is ungated (gated=false) and available for immediate download. Recommend: (1) include Apache 2.0 text in any product shipping this model; (2) document any modifications; (3) review Qwen's separate commercial terms if using via Qwen API or managed services. No liability exclusions beyond standard Apache 2.0; evaluate your risk tolerance.
DEV.co evaluation signals
Editorial assessment — not user reviews. Directional, with an explicit confidence level.
| Signal | Assessment |
|---|---|
| Maintenance | Active |
| Documentation | Strong |
| License clarity | Clear |
| Deployment complexity | Moderate |
| DEV.co fit | Strong |
| Assessment confidence | High |
No explicit security audit or adversarial robustness testing described in the card. Model is a language model; standard LLM risks apply: prompt injection, jailbreaking, hallucination, and environment-prediction inaccuracy under adversarial inputs. The model's specialized world-model task (next-state prediction) means hallucinated environment states could mislead agents. For agentic deployment: validate predicted states against real or semi-real environments; implement agent-level safeguards (e.g., action validation, rollback). No disclaimer of safety guarantees. Training data disclaimer: 'No outputs from external API services are included.' Recommend: treat predictions as simulations not ground truth; security-sensitive agent tasks need additional validation layers.
Alternatives to consider
GPT-5.4 or Claude Opus 4.8
Closed-source, hosted APIs. GPT-5.4 scores 58.25 on AgentWorldBench vs. Qwen-AgentWorld 56.39. Better for immediate production if cost/latency acceptable. No self-hosting or fine-tuning. Trade-off: vendor lock-in, data privacy concerns, higher per-query cost.
Qwen3.5-397B-A17B (larger variant)
Same training pipeline as Qwen-AgentWorld but 397B parameters (17B active). Scores 58.71 on AgentWorldBench vs. 56.39 for 35B variant. Better accuracy if hardware/cost permits. Trade-off: requires more VRAM, slower inference, higher resource cost.
Open-source base models (Llama 3.1, Mistral) + custom fine-tuning
Lower baseline cost and smaller footprint. Enables full control over training data and domain specialization. Trade-off: requires significant engineering effort for world-model training pipeline (CPT + SFT + RL), no pre-built agent-environment knowledge, longer time-to-production.
Ship Qwen-AgentWorld-35B-A3B with senior software developers
Download the model from Hugging Face (ungated, Apache 2.0). Use vLLM or SGLang for fast inference with OpenAI-compatible APIs. Consult domain-specific system prompts in the GitHub repository and benchmark performance on AgentWorldBench to validate fit for your agentic tasks.
Talk to DEV.coRelated open-source tools
Surfaced by semantic similarity across the DEV.co open-source index.
Related on DEV.co
Explore the category and the services that help you build with it.
Qwen-AgentWorld-35B-A3B FAQ
Can I use this model commercially in a production SaaS product?
What GPU hardware do I need to run this model?
How does this compare to fine-tuning a base model myself?
What is the difference between Qwen-AgentWorld-35B-A3B and the base model?
Software developers & web developers for hire
DEV.co helps companies turn open-source tools like Qwen-AgentWorld-35B-A3B into production software. Our software development services cover the full lifecycle — architecture, web development, integration, and maintenance — delivered by software developers and web developers who ship. Engage our software development agency to implement or customize it for your open-source llms stack.
Ready to Deploy Qwen-AgentWorld for Agent Simulation?
Download the model from Hugging Face (ungated, Apache 2.0). Use vLLM or SGLang for fast inference with OpenAI-compatible APIs. Consult domain-specific system prompts in the GitHub repository and benchmark performance on AgentWorldBench to validate fit for your agentic tasks.