---

base_model: microsoft/Phi-4-mini-instruct

library_name: transformers

license: mit

pipeline_tag: text-generation

tags:

• gguf
• llama-cpp
• robotics
• drones
• uav
• mavlink
• px4
• ardupilot
• edge-ai
• local-llm
• q4_k_m

---

OpenZero SkyPilot 3.8B GGUF

!OpenZero SkyPilot hero

Private research release — a verified local GGUF for simulator-first aerial logistics and mission-proposal research. It is not a certified flight controller or deployment-ready autonomy system.

OpenZero SkyPilot 3.8B is an experimental mission-level aerial-vehicle proposal model derived from the pinned microsoft/Phi-4-mini-instruct revision cfbefacb99257ffa30c83adab238a50856ac3083.

The model is intended to produce typed, reviewable mission proposals for integration research around MAVLink-compatible simulators such as PX4 SITL and ArduPilot SITL. Stabilization, geofencing, collision avoidance, flight envelopes, return-to-home, emergency handling and operator authority remain with deterministic systems outside the LLM.

Verified GGUF artifact

| Field | Verified value |

|---|---|

| File | OpenZero-SkyPilot-3.8B-Q4_K_M.gguf |

| Quantization | Q4_K_M |

| Size | 2,493,840,096 bytes |

| SHA-256 | 10e3dad8684d74cd62162c497e27bf7c0ed62a22818d16795b163330e04dffd7 |

| llama.cpp build | b10451 |

| llama.cpp commit | 10bf611e533d81f739128304991c5e133c6aebd8 |

| Runtime gate | Passed: real GGUF inference, 8 generated tokens |

| Observed CPU generation | 5.6 tokens/s in the recorded gate |

The machine-readable verification record is included as OpenZero-SkyPilot-3.8B-GGUF-Evidence.json.

Research use cases

• simulator-first aerial logistics and approved package-handling proposals
• waypoint, altitude and return-to-home mission-envelope testing
• MAVLink/PX4/ArduPilot adapter conformance research
• infrastructure inspection and environmental survey planning
• fault injection, degraded-link and abort-policy evaluation
• typed mission proposals for independent validation and operator approval

SkyPilot is a language-level mission proposer. It does not replace a flight controller, state estimator, collision-avoidance system or qualified operator.

Training and provenance

• Base: microsoft/Phi-4-mini-instruct, exact revision pinned above.
• Finite two-step QLoRA smoke run on a Tesla T4.
• 32 deterministic locally authored training cases.
• 6 disjoint validation cases.
• Recorded training loss: 4.62113618850708.
• Adapter SHA-256: 1296b817cd2d2063bb23f348361e9a77c604e9742c10b18eb43b61d62fba6f38.
• Excluded: Fusion data, teacher outputs, locked evaluation, rejected data, Ministral data and failed Gemma-31B checkpoints.

This deliberately small smoke dataset establishes pipeline integrity and artifact reproducibility. It does not establish broad aerial competence, safety certification or superiority over the exact base model.

Run with llama.cpp

llama-cli \
  -m OpenZero-SkyPilot-3.8B-Q4_K_M.gguf \
  -cnv \
  -p "Propose a simulator-only inspection mission as typed JSON."

Treat every output as untrusted proposed data. Validate it against a signed capability manifest, schema, geofence, vehicle limits, current telemetry and explicit operator authority before any simulator adapter accepts it.

Integration boundary

operator intent
    -> SkyPilot typed mission proposal
    -> schema and capability validation
    -> policy, geofence and fault checks
    -> operator approval
    -> MAVLink simulator adapter
    -> PX4/ArduPilot SITL deterministic control

Approved package pickup or delivery research must use declared benign payload constraints, verified pickup/drop zones and operator-controlled abort behavior.

Explicit exclusions

SkyPilot is not designed for weapons or harmful payloads, person targeting or pursuit, threat engagement, evasion, interference, unauthorised control, raw motor/actuator access, flight-controller replacement or safety-system bypass.

Status and claims

experimental research artifact. No MOD, UKRI, OpenAI, Microsoft or other institutional approval, certification, procurement or endorsement is claimed.