DEV.co
Open-Source LLM · unsloth

Qwen-AgentWorld-35B-A3B-GGUF

Qwen-AgentWorld-35B-A3B is a 35-billion-parameter language model specialized for simulating agent environments across seven domains: tool calling, search, terminal, software engineering, Android, web, and OS interaction. It uses a GGUF-quantized format for efficient deployment and is designed to predict environment state changes given agent actions, functioning as a 'world model' rather than a general-purpose chatbot. Licensed under Apache 2.0 with no gating restrictions.

Source: HuggingFace — huggingface.co/unsloth/Qwen-AgentWorld-35B-A3B-GGUF
Unknown
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
377.3k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developerunsloth
ParametersUnknown
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads377.3k
Likes156
Last updated2026-06-25
Sourceunsloth/Qwen-AgentWorld-35B-A3B-GGUF

What Qwen-AgentWorld-35B-A3B-GGUF is

Native language world model (Qwen3.5 base, 35B total parameters with 3B activated via MoE) trained through continual pre-training, supervised fine-tuning, and reinforcement learning (GSPO). Features 262,144-token context window, mixed Gated DeltaNet and attention layers, 256 MoE experts (8 routed + 1 shared), and is quantized in GGUF format by Unsloth. Comparable to vLLM and SGLang serving frameworks. No external API outputs in training.

Quickstart

Run Qwen-AgentWorld-35B-A3B-GGUF locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="unsloth/Qwen-AgentWorld-35B-A3B-GGUF")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.

Deployment

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

Autonomous Agent Development

Simulate multi-step agentic workflows across terminal, web, Android, and software engineering tasks. The model's native world-model training makes it suitable for planning agents that need to predict environment responses without calling real services.

Environment Simulation for Agent Testing

Use as a synthetic environment backend for agent development and debugging. Reduces dependency on real service APIs and enables controlled perturbation injection (fictional scenarios, edge cases). Particularly strong for SWE, terminal, and search domains.

Self-Hosted Multi-Domain Agent Backbone

Deploy on-premise or in private infrastructure via vLLM/SGLang. The Apache 2.0 license and GGUF quantization enable full control over model execution, suitable for regulated environments needing deterministic agent simulation.

Running & fine-tuning it

Minimum 8–16 GB VRAM per GPU for single-device inference (GGUF quantization reduces from ~70 GB full precision). For production serving with throughput: 4× A100/H100 or 8× RTX 6000 Ada recommended. Context length (262K tokens) requires proportional memory; reducing to 128K tokens saves ~50% memory. Tensor parallelism (tp-size 4 in examples) distributes load; estimate ~14 GB per GPU for tp-size=4 deployment.

Unknown. Card does not disclose LoRA/QLoRA compatibility, parameter count specifics for adapter training, or recommended fine-tuning frameworks. GGUF format may complicate direct fine-tuning workflows (typically requires conversion back to full precision or LoRA adapter layer integration). Recommend consulting Unsloth documentation or Qwen GitHub for adapter layer support.

When to avoid it — and what to weigh

  • Real-Time, Single-Digit-Millisecond Inference — 35B parameters (even with MoE sparsity) require substantial compute. Unsloth GGUF format helps, but latency will exceed 100ms for typical hardware; not suitable for sub-100ms SLA workloads.
  • Factual Knowledge Cutoff or Real-Time Information — Model is optimized for environment simulation, not knowledge retrieval. Will underperform on open-ended QA, current events, or domains where ground-truth facts matter more than plausible trajectory prediction.
  • Fully Autonomous Deployment Without Verification — Model-generated predictions (terminal output, webpage states) are simulated and may diverge from ground truth. Requires human-in-the-loop validation or integration with real environment feedback for production agentic systems.
  • Consumer-Grade or Severely Memory-Constrained Hardware — Even quantized, 35B requires minimum ~8–16 GB VRAM with tensor parallelism or dynamic batching. Estimate ~14 GB VRAM for single-GPU inference; recommend 4× H100 or equivalent for production throughput.

License & commercial use

Apache License 2.0 (apache-2.0). Permissive OSI-approved license allowing use, modification, and distribution with attribution. No restrictions on source form or derivative works. No viral copyleft.

Apache 2.0 is a permissive OSI license that explicitly permits commercial use, including for proprietary derivative works. You may build commercial products using Qwen-AgentWorld-35B-A3B, train on proprietary data, and redistribute. No royalties or special licensing required. No gating; model is publicly available. However, verify compliance with Qwen's parent org (Alibaba) usage policies and any downstream service terms (vLLM, SGLang, Hugging Face) if deploying via those platforms.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationStrong
License clarityClear
Deployment complexityModerate
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

Card states 'No outputs from external API services included in training.' This mitigates data-poisoning risk. GGUF quantization is a binary format; verify checksum integrity when downloading. Model operates on local inputs and generates predictions; no telemetry or API callouts disclosed. If deploying via cloud (vLLM/SGLang API), apply standard API security (authentication, rate-limiting, input validation). Multi-turn reasoning with long context may increase token-injection risk in multi-user environments; isolate user contexts. No third-party security audit mentioned; recommend custom red-teaming for production agentic workflows.

Alternatives to consider

GPT-4o or Claude Opus 4

Closed-source alternatives with higher AgentWorldBench scores (58.25, 56.59 vs. 56.39 for 35B model). Trade off cost, API dependency, and commercial terms for stronger multi-domain performance and real-time fact access.

Qwen-AgentWorld-397B-A17B

Larger sibling (397B, 17B activated) from same family; 58.71 benchmark score vs. 56.39. Higher compute cost but stronger performance, especially in search and SWE. Same license and training approach.

DeepSeek-V4-Pro or other open-weight MoE models

Comparable parameter efficiency and open-source availability. Lower AgentWorldBench scores in world-model tasks but may offer stronger general reasoning. Evaluate trade-off between domain specialization and generality.

Software development agency

Ship Qwen-AgentWorld-35B-A3B-GGUF with senior software developers

Download Qwen-AgentWorld-35B-A3B GGUF from Hugging Face and deploy via vLLM or SGLang. Consult the technical report and GitHub repository for domain-specific system prompts and integration examples.

Talk to DEV.co

Related 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-GGUF FAQ

Can I use Qwen-AgentWorld-35B-A3B in a commercial product?
Yes. Apache 2.0 license permits unrestricted commercial use. You may train proprietary models on it, sell products, and modify as needed. No royalties or special licensing. Verify integration policies if serving via third-party platforms (vLLM, SGLang, Hugging Face).
What GPU hardware do I need to run this model?
Single GPU: minimum 8–16 GB VRAM (estimate ~14 GB for GGUF quantization). Production: 4× A100/H100 or equivalent, with tensor-parallelism (tp-size 4). Context length (default 262K tokens) scales memory linearly; reduce to 128K if OOM occurs, though that limits multi-turn reasoning.
How does Qwen-AgentWorld compare to general-purpose LLMs like GPT or Claude?
Specialized for environment simulation (predicting next state from agent actions). AgentWorldBench shows Qwen-AgentWorld-35B (56.39) scores below GPT-4 (58.25) and Claude Opus (56.59) but is open-source, self-hostable, and designed for agentic workflows. Larger 397B variant (58.71) approaches closed-source performance. General QA/facts: general LLMs are stronger.
Can I fine-tune this model?
Unknown. Card does not specify LoRA/QLoRA support or fine-tuning procedures for GGUF format. Likely requires conversion to full precision or adapter-layer integration. Consult Unsloth and Qwen GitHub repositories for guidance; standard Transformers fine-tuning may not directly apply to quantized weights.

Software developers & web developers for hire

Adopting Qwen-AgentWorld-35B-A3B-GGUF is usually one piece of a larger software development effort. As a software development agency, DEV.co provides software development services and web development expertise — pairing senior software developers and web developers with your team to design, build, and operate open-source llms software in production.

Start Simulating Agent Environments Today

Download Qwen-AgentWorld-35B-A3B GGUF from Hugging Face and deploy via vLLM or SGLang. Consult the technical report and GitHub repository for domain-specific system prompts and integration examples.