---

license: apache-2.0

base_model: mistralai/Ministral-3-8B-Instruct-2512-BF16

pipeline_tag: text-generation

inference: false

tags:

• gguf
• llama-cpp
• ministral
• openzero
• agent-zero
• local-ai
• agentic
• tool-use
• coding
• research
• q5-k-m

language:

• en

---

OpenZero Ministral 3 8B Runtime Agent GGUF

Extensive testing has been carried out to create this LLM for OpenZero and other agentic software. 1 billion tokens used in tests gpt 5.6 codex to ensure this LLM is high quality!

10 LMM's were tested this LLM is the final result as it passed all agentic tests, high effort went into this LLM

!OpenZero Ministral 8B Runtime Agent

A local GGUF for llama.cpp, based on Ministral 3 8B Instruct, packaged for OpenZero / Agent Zero workflows involving coding, research, structured tool use, privacy-aware operation and evidence-bounded answers. The release contains one verified Q5_K_M model for practical CPU deployment.

> Provenance: upstream model weights are unchanged. This is a runtime-template edition, not a weight fine-tune and not a cross-model merge.

Download

| File | Quantization | Bytes | SHA-256 | Use |

|---|---:|---:|---|---|

| OpenZero-Ministral3-8B-Runtime-Agent-Q5_K_M.gguf | Q5_K_M | 6,058,748,288 | e9aba29e5465164933d334215c2e8d5d9edddfd5caf71ecce6c5f811ceb11d9e | Recommended local CPU balance |

hf download shafire/OpenZero-Ministral3-8B-Runtime-Agent-GGUF OpenZero-Ministral3-8B-Runtime-Agent-Q5_K_M.gguf --local-dir .

Run with llama.cpp

llama-cli -m OpenZero-Ministral3-8B-Runtime-Agent-Q5_K_M.gguf -c 4096 -n 512 -ngl 0

For constrained RAM, start with -c 2048. GPU offload can be enabled by increasing -ngl where supported.

What changed from upstream

• Preserved the official upstream weights exactly.
• Patched and embedded the standalone chat_template.jinja as a GGUF runtime template.
• Added OpenZero / Agent Zero identity and one-operator-tool-per-turn guidance.
• Added evidence, uncertainty, privacy, provenance and credential-protection guidance.
• Did not train on private data, merge adapters, or alter model tensors.

The embedded guidance expects an OpenZero-compatible orchestrator. The model can emit a structured tool request, but the model itself cannot browse, read files, execute commands, rewrite applications or act autonomously. The host runtime must validate permissions, arguments and results.

Reproducibility

• Upstream: mistralai/Ministral-3-8B-Instruct-2512-BF16
• Immutable upstream revision: f6fae9795746f63c9be8344932f01275f3c63734
• Upstream licence: Apache-2.0
• llama.cpp release: b10333
• llama.cpp source commit: 08659901c43b51de735740f1cf61bb82fbe0c4e4
• Official Linux binary archive SHA-256: 936ce04d98abe2a977e9dd2ff92659bb96947e136acee8f2bc3e21d8eaebbf23
• Conversion path: upstream BF16 checkpoint to F16 GGUF, then Q5_K_M quantization.
• conversion_report.json and SHA256SUMS.txt are included.

Validation

The GGUF magic, byte size, SHA-256 and embedded template were verified. It also passed a bounded one-token CPU text load test using llama.cpp b10333:

llama-cli -m OpenZero-Ministral3-8B-Runtime-Agent-Q5_K_M.gguf -p "Hello" -c 128 -n 1 -ngl 0 --no-conversation --single-turn --simple-io --no-warmup

This proves the artifact loads and generates text; it is not a general capability benchmark.

Intended use

• Local OpenZero and Agent Zero experimentation
• Coding and research assistance
• Structured tool-call orchestration
• Private/offline CPU inference
• Reproducible GGUF and llama.cpp testing

Limitations

• Runtime instructions improve consistency but do not guarantee compliance.
• Validate tool calls and generated code before execution.
• Current facts require an external retrieval tool; the model should not invent them.
• Q5_K_M is lossy relative to upstream BF16.
• English is the primary tested language.
• This release is text-only; no vision projector is included.
• Do not provide secrets in prompts or logs.

Attribution and licence

Derived from Mistral AI's Ministral 3 8B Instruct under Apache-2.0. This repository does not imply endorsement by Mistral AI.

Upstream: https://huggingface.co/mistralai/Ministral-3-8B-Instruct-2512-BF16

Collection: https://huggingface.co/collections/shafire/agentic-gguf-models