shafire/OpenZero-MeshBridge-3.8B-GGUF overview
OpenZero MeshBridge 3.8B GGUF OpenZero MeshBridge hero OpenZero MeshBridge 3.8B hero.png Private, access controlled research release for authorised simulator i…
Runs locally from ~2.32 GB disk (4 GB VRAM class GPUs with llama.cpp / guIDE).
Repository Files & Downloads
| File | Type | Quantization | Size | Link |
|---|---|---|---|---|
| OpenZero-MeshBridge-3.8B-Q4_K_M.gguf | GGUF | Q4_K_M | 2.32 GB | Download |
Model Details
| Model ID | shafire/OpenZero-MeshBridge-3.8B-GGUF |
|---|---|
| Author | shafire |
| Pipeline | text-generation |
| License | mit |
| Base model | microsoft/Phi-4-mini-instruct |
| Last modified | 2026-10-07T02:10:45.000Z |
Model README
---
base_model: microsoft/Phi-4-mini-instruct
library_name: transformers
license: mit
pipeline_tag: text-generation
tags:
- gguf
- llama-cpp
- robotics
- ros2
- mavlink
- interoperability
- edge-ai
- local-llm
- q4_k_m
---
OpenZero MeshBridge 3.8B GGUF
Private, access-controlled research release for authorised simulator interoperability. It is not a universal takeover system, certified controller or deployment-ready autonomy product.
MeshBridge is an experimental mission-level supervisory proposal model derived from the pinned microsoft/Phi-4-mini-instruct revision cfbefacb99257ffa30c83adab238a50856ac3083. It explores auditable onboarding of an explicitly authorised simulator platform through signed capability manifests, known adapters, deterministic conformance tests and time-bounded operator handoff.
Verified artifact
| Field | Verified value |
|---|---|
| File | OpenZero-MeshBridge-3.8B-Q4_K_M.gguf |
| Quantization | Q4_K_M |
| Bytes | 2,493,840,128 |
| SHA-256 | 439841f6b924055bfd296373605237e2557421f3f5438a05dc13b0b44b3406eb |
| llama.cpp | b10451, commit 10bf611e533d81f739128304991c5e133c6aebd8 |
| Runtime gate | Passed: real llama.cpp inference, 8 generated tokens |
| Observed CPU generation | 5.8 tokens/s in the recorded gate |
Machine-readable evidence is included as OpenZero-MeshBridge-3.8B-GGUF-Evidence.json.
Authorised commissioning sequence
DISCOVER_CAPABILITIES
-> VERIFY_SIGNED_MANIFEST
-> BIND_SIMULATOR_ADAPTER
-> RUN_CONFORMANCE_SUITE
-> REQUEST_CONTROL_HANDOFF
-> ASSUME_AUTHORISED_SUPERVISORY_CONTROL
-> RELINQUISH_CONTROL
Unknown interfaces fail closed. A session is limited to the exact signed platform, declared capabilities, permitted scope and validity window. Deterministic controllers retain real-time limits, emergency stop and actuation.
Evidence and provenance
- Exact pinned base revision recorded above.
- 32 deterministic locally authored training cases and 6 disjoint held-out cases.
- Two-step finite QLoRA smoke run passed on a Tesla T4; loss
5.029376983642578. - Adapter SHA-256:
9dd17d11de0521c97b6b0ec76bc2e2c657cdb89abc79427a23b5b8f24f363b3e. - Held-out exact-base loss:
5.684460004170735. - Held-out adapter loss:
5.266995271046956. - Improvement:
0.4174647331237793lower loss on the immutable six-case set. - Fusion, teacher outputs, locked evaluation, rejected data, Ministral data and failed Gemma-31B checkpoints were excluded.
The small evaluation demonstrates a reproducible pipeline-level improvement, not broad robotics competence, universal compatibility or safety certification.
Intended research uses
- signed capability-manifest parsing and validation
- known ROS 2/Nav2, ros2_control and MAVLink simulator adapter selection
- stop, hold, limits, stale telemetry, link loss, geofence and rollback conformance testing
- operator-reviewed mission-level handoff and automatic relinquishment
- approved benign logistics simulations with declared package constraints
- passive emergency-assistance proposals without threat engagement or force
Run locally
llama-cli \
-m OpenZero-MeshBridge-3.8B-Q4_K_M.gguf \
-cnv \
-p "Inspect this signed simulator capability manifest and propose the next commissioning stage as typed JSON."
Treat all output as untrusted proposed data. Independently validate schemas, signatures, scope, telemetry freshness, platform limits and operator authority.
Explicit exclusions
MeshBridge is not designed for arbitrary or unauthorised takeover, unknown interfaces, raw actuator or flight-control access, weapons or harmful payloads, person targeting or pursuit, threat engagement, force, restraint, evasion, interference or safety-controller bypass.
Status
experimental research artifact. No MOD, UKRI, OpenAI, Microsoft or other institutional approval, certification, procurement, deployment or endorsement is claimed.
Run shafire/OpenZero-MeshBridge-3.8B-GGUF with guIDE
Download guIDE — the AI-native code editor with local LLM inference and 69 built-in tools.
Source: Hugging Face · Compare models