greppy-qwen35-mtp-q4km

greppy Qwen3.5-0.8B — function-purpose / MTP — Q4_K_M GGUF

A full-parameter fine-tune of Qwen3.5-0.8B trained in-house by metricspace for greppy, the local code-navigation tool for coding agents. greppy embeds this file in its binary and runs it on device (native Rust + vendored Metal/CUDA kernels — no llama.cpp runtime, Python, or model server) to write the one-line purpose hint shown under each function signature that greppy returns (brief, who-calls, semantic-search, …).

This is not a stock checkpoint. It is greppy's own model, distilled specifically for code-navigation summaries.

• Model id: greppy-qwen35-function-purpose-2026-07-11
• Base: Qwen/Qwen3.5-0.8B
• Training: full-parameter supervised fine-tuning with FP32 master weights, distilled from a larger teacher toward one-line function-purpose hints.
• MTP head: trained for multi-token-prediction (speculative-draft depth); offline MTP acceptance ≈ 79.5 %, eval-loss ≈ 1.34 on the promoted checkpoint (model-qwen35-nano-20260716-0749).
• Format: Q4_K_M GGUF (plus tokenizer.json).

License

Apache-2.0 (inherited from Qwen3.5-0.8B). Attribution, the record of modifications, and full training/data provenance are included:

• QWEN3.5-APACHE-2.0.txt
• QWEN3.5-MODIFICATIONS.txt
• QWEN3.5-PROVENANCE.json
• QWEN3.5-TRAINING-DATA-MANIFEST.json

Intended use

Purpose-hint generation and short code summaries inside greppy. The hints are navigation aids, not authority: greppy attaches them to the exact source signature, and callers are expected to read the returned source before acting.