shafire/OpenZero-Fusion-Qwen3-4B-Agentic-GGUF

OpenZero Fusion Qwen3-4B Agentic — Standalone GGUF

THIS MODEL WORKS BUT HAS OUTPUT ERROR ISSUES.NOT RECOMMENDED FOR PRODUCTION. ONE FILE. TWO SPECIALISTS DISTILLED INTO ONE LOCAL AGENT.

[image]

OpenZero Fusion Qwen3-4B Agentic is a single-architecture Qwen3-4B student trained from the curated OpenZero corpus plus filtered outputs from the completed Qwen3-1.7B and Gemma4-E2B specialists. It is knowledge distillation—not an invalid direct weight merge between unrelated architectures. Gemma 4 and Qwen 3 into 1 LLM made for CPU. Experimental model results may vary for this model.

What is included

• —Standalone model: yes
• —Separate adapter required: no
• —Separate base model required: no
• —Base architecture: Qwen/Qwen3-4B
• —Training mix: 2,606 curated examples + 379 filtered two-teacher examples = 2,985 training rows
• —Held-out evaluation: 137 rows, never used for training
• —Final held-out loss: 1.643606
• —Held-out token accuracy: 0.788742
• —GGUF conversion and CPU load test: passed for all three files with llama.cpp

Run with llama.cpp

hf download shafire/OpenZero-Fusion-Qwen3-4B-Agentic-GGUF OpenZero-Fusion-Qwen3-4B-Agentic-Q4_K_M.gguf --local-dir .

llama-cli -m OpenZero-Fusion-Qwen3-4B-Agentic-Q4_K_M.gguf --jinja -c 8192 -t 8 --temp 0.6 --top-p 0.95

For a local OpenAI-compatible endpoint:

llama-server -m OpenZero-Fusion-Qwen3-4B-Agentic-Q4_K_M.gguf --jinja -c 8192 -t 8 --host 127.0.0.1 --port 8080

Why Fusion exists

The small Qwen and Gemma specialists cannot safely be tensor-merged: their architectures and tokenizers differ. Fusion instead trains one compatible Qwen3-4B student using useful outputs from both teachers, producing a single deployable GGUF.

Verified release

• —Full QLoRA training: 94/94 steps, one epoch
• —Train loss: 2.158780
• —Held-out evaluation loss: 1.643606
• —Held-out token accuracy: 0.788742
• —Q4_K_M — 2,497,280,288 bytes — SHA-256 f00ad7bb04cb88c6861070e4b6176c9ee208030ae91ea9bb9ebeece7ee9ae415
• —Q8_0 — 4,280,404,768 bytes — SHA-256 8430bcfe6704c53cfa0f8d3e30d141d78730e1d47bcd80515413ec24a090317c
• —F16 — 8,051,284,768 bytes — SHA-256 29069c52929b50681d08cfc2564cf7e443b3b0c286fbf6b8ad1d285f11d11196

Provenance and reproducibility

The student used 2,606 original OpenZero instruction rows plus 379 filtered teacher-response rows, for 2,985 training rows total. Teacher generation began with 192 balanced prompts across general, coding, research and agent/tool categories; both specialists produced a candidate for each prompt. A separate 137-row held-out set was never used for training. The completed V6 adapter was merged only into Qwen/Qwen3-4B revision 1cfa9a7208912126459214e8b04321603b3df60c. F16 was converted with llama.cpp b10333 / commit 08659901c43b51de735740f1cf61bb82fbe0c4e4; Q80 and Q4K_M were independently quantized from that F16 source. Every file passed a bounded one-shot CPU text load test (-c 128 -n 1 -ngl 0 --no-conversation --single-turn --simple-io --no-warmup) and remote LFS/Xet SHA-256 verification.

This distillation transfers examples, not tensors. It does not make a 4B model equivalent to the sum of two architectures, and the reported held-out metrics are not broad benchmark claims. Tool calls are text emitted for an external runtime to validate and execute.

This model is an independent fine-tune based on Qwen and is not affiliated with or endorsed by Qwen. The Qwen base is Apache-2.0. OpenZero training materials and resulting community release are subject to the OpenZero Community Source terms; do not describe this release as OSI-approved open source. Review both upstream and OpenZero terms before redistribution or commercial use.