---

language:

• en

library_name: llama.cpp

pipeline_tag: text-generation

inference: false

tags:

• gguf
• gemma
• local-ai
• agentic-ai
• coding-assistant
• research-assistant
• tool-use
• openzero
• ollama
• llama-cpp
• quantized

base_model: google/gemma-4-E2B-it

---

OpenZero Gemma4 E2B Agentic — Standalone GGUF

ONE FILE. NO ADAPTER. NO BASE-MODEL HUNT.

!OpenZero Gemma4 E2B Agentic

OpenZero Gemma4 E2B Agentic is a small local GGUF research, coding and tool-workflow specialist. Choose the verified Q4_K_M, Q8_0 or F16 build; no PEFT archive or separate base model is required to run it.

What is included

| File | Purpose |

|---|---|

| OpenZero-Gemma4-E2B-Agentic-Q4_K_M.gguf | Recommended balance of size and quality |

| OpenZero-Gemma4-E2B-Agentic-Q8_0.gguf | Higher fidelity, larger download |

| OpenZero-Gemma4-E2B-Agentic-F16.gguf | Reference full-precision GGUF |

• Standalone model: yes
• Separate adapter required: no
• Separate base model required: no
• Base architecture: google/gemma-4-E2B-it
• Fine-tuning: 2,606 training examples; 137 held-out evaluation examples
• Final held-out loss: 1.520900
• Held-out token accuracy: 0.657464
• GGUF conversion and CPU load test: passed for all three files with llama.cpp
• Modality in this repository: text-only; no mmproj vision/audio projector is included

Run with llama.cpp

hf download shafire/OpenZero-Gemma4-E2B-Agentic-GGUF OpenZero-Gemma4-E2B-Agentic-Q4_K_M.gguf --local-dir .

llama-cli -m OpenZero-Gemma4-E2B-Agentic-Q4_K_M.gguf --jinja -c 8192 -t 8 --temp 0.6 --top-p 0.95

Positioning

Built for private local inference, practical code work, evidence-aware research and agent runtimes. Tool execution belongs to the surrounding runtime; validate outputs before acting on them.

Verified release

• Training adapter: QLoRA specialist run, 82/82 steps
• Train loss: 2.261954
• Held-out evaluation loss: 1.520900
• Held-out token accuracy: 0.657464
• Q4_K_M — 3,416,119,872 bytes — SHA-256 9a7e717d13208526782c5fa5074bbcd7f445c6ac720e88d0e0e3ae6d87ebdeef
• Q8_0 — 4,947,414,592 bytes — SHA-256 b612e7a6ffc2b9707f3b68e7e94fc8f8764e0975fefa4d1256717356f18af19f
• F16 — 9,273,527,872 bytes — SHA-256 51e25487e2193df69b9271ae8ad7bdaffb6197a861b65f6e27c70e2a556fbbb0

Provenance and reproducibility

The V12 LoRA adapter was trained for 82 steps (one epoch) on 2,606 OpenZero instruction examples. A separate 137-row held-out set was used only for final evaluation. Training used unsloth/gemma-4-E2B-it-unsloth-bnb-4bit; the adapter was merged into google/gemma-4-E2B-it revision 3e22461f65e89153144f8adb70e3b8c2cc9845a7. The merged model was converted with llama.cpp b10333 / commit 08659901c43b51de735740f1cf61bb82fbe0c4e4; Q8_0 and Q4_K_M were independently quantized from F16. A narrowly asserted converter compatibility shim derived global_head_dim=512 from the preserved per-layer Gemma configuration. 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 size/LFS SHA verification.

This is a practical specialist fine-tune, not a claim of superiority over the base model. Tool calls are text emitted for an agent runtime to validate and execute. This repository does not include Gemma's multimodal projector.

This model is an independent fine-tune based on Gemma and is not affiliated with or endorsed by Google. Preserve Google Gemma attribution and applicable upstream terms. 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.