---

license: apache-2.0

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

tags: [gguf, ollama, llama.cpp, agentic-coding, reasoning, world-model, qwen35moe, moe]

language: [en]

---

!banner

Qwen-AgentWorld-35B-A3B — IQ2_M GGUF (first-mover)

> Qwen's AgentWorld 35B-total / 3B-active hybrid MoE (linear-attention + sparse experts), quantized to a ~11.6 GB IQ2_M that runs on a 16 GB machine and autonomously solves real SWE-bench Verified bugs. This is Qwen's model — ours is the quant (imatrix) + the eval + the configurable-context Modelfiles. No GGUF existed yet; this is the first.

Heads-up — brand-new architecture: AgentWorld is qwen35moe, a hybrid that mixes linear-attention (gated delta-net + conv) layers with full-attention layers, on top of a 256-expert MoE (8 active). It needs a recent llama.cpp / Ollama with qwen35moe runtime support — older builds (and Ollama ≤ 0.20.x at time of writing) cannot run it. Build llama.cpp from master.

📦 Which quant should I pick?

| file | bits | size | pick this if… |

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

| AgentWorld-35B-A3B-IQ2_M.gguf | ~2.7 | ~11.6 GB | the one in this repo — smallest that's still sharp (imatrix-guided). Runs in ~13–16 GB. |

<!-- I-quants (IQ*) are smaller than K-quants at the same quality but need an imatrix (shipped here as AgentWorld-35B.imatrix) and a bit more CPU. Because only 3B params are active per token, even this 35B-total model is fast. -->

🧮 Will it fit? (RAM/VRAM cheat-sheet)

Total memory ≈ ~11.6 GB weights + KV-cache (grows with context). Only 3B params are active per token, so it's fast for its size.

| you have | context you can run |

|---|---|

| 16 GB | ~8–32K |

| 24 GB | ~64–128K |

| 32 GB+ | up to native 256K |

🧠 Context size (native 262144 / 256K)

Configurable — Modelfiles at 8K / 32K / 128K / native are included; see CONTEXT.md.

🚀 How to run it

# Build a recent llama.cpp first (master has qwen35moe support):
#   git clone https://github.com/ggml-org/llama.cpp && cmake -B build -DGGML_METAL=ON && cmake --build build -j

# llama.cpp (OpenAI-compatible server, native tool-calls via --jinja):
llama-server -m AgentWorld-35B-A3B-IQ2_M.gguf -c 32768 --jinja -ngl 99

# Ollama (recent build that supports qwen35moe):
ollama create agentworld-32k -f Modelfile.32768 && ollama run agentworld-32k

Recommended sampling: temperature 1.0, top_p 0.95 (this is a reasoning model — low temp causes repetition loops). It emits a <think> reasoning channel; --jinja parses it into reasoning_content and tool calls into native tool_calls.

Agentic use: it makes native OpenAI tool calls out of the box with --jinja (no <function=>/text-injection workaround needed). In a Claude-Code-style harness it also reliably emits raw-JSON {"name":…, "arguments":…} tool calls.

⚠️ Good to know

• Strengths: agentic coding / tool use — solves real SWE-bench Verified bugs autonomously; fast (3B active); coherent reasoning at IQ2.
• Limits: at IQ2_M it occasionally rambles on a single long turn in very long agentic context (cap max_tokens ~8K to bound it). It's a world-model/agent model (the upstream is multimodal; this GGUF is text-only — vision is not converted).
• Needs a current runtime (see the architecture note above).

📊 Evaluation methodology (MANDATORY — never lose how it was measured)

• Score: 1/3 resolved on a 3-instance swe-mix anchor smoke (see below). The confirmed real solve: django__django-11099 (is_resolved: True, 1/1 tests, 37 tool calls, autonomous). django__django-11133 attempted but not resolved (0/1); astropy__astropy-14369 (a "1-4 hour" hard task) timed out under the 25-min watchdog. (A full 30-instance swe-mix run was not completed for this first-mover release; this is an honest agentic smoke, not the full probe.)
• Exact sample (smoke): django__django-11099 (resolved), django__django-11133 (0/1), astropy__astropy-14369 (timeout) — anchors from the swe-mix probe (full 30-instance spec in swe_mix.json: 6 continuity anchors + 24 discriminating, 10 repos, SWE-bench Verified).
• Harness: Claude Code + agent-bridge.js, terminal-bench / Docker, served by llama-server (master build, Metal) on an OpenAI /v1 endpoint.
• Sampling: temp 1.0, top_p 0.95. Tool format: native OpenAI tool_calls (via --jinja) → bridge raw-JSON.
• Quant: IQ2_M, imatrix on a ~550 KB code+general corpus (273 chunks, final PPL 1.56), ~11.6 GB. Quantized from a Q8_0 round-trip (--allow-requantize).
• Date: 2026-06-28. Caveat: small/hard local smoke, relative signal — NOT 1:1 with the official SWE-bench leaderboard.

🗒️ Changelog

• 2026-06-28 v1: initial IQ2_M release (first GGUF of AgentWorld-35B-A3B).

📚 Credit & license

Model, weights, training: © Qwen / Alibaba — Qwen/Qwen-AgentWorld-35B-A3B. Quant (IQ2_M + imatrix) + context Modelfiles + eval: KikoCis. apache-2.0 (same as upstream). No weights modified beyond quantization; text-only (vision tower not converted).