Verdict 4 500M Quill
50-iteration fused LoRA SFT on Qwen2.5-Coder-0.5B-Instruct (~1.34 GiB peak RSS at 32K context) for ultra-constrained 4–8 GB RAM laptops and fast local diagnostics.
Lineage, Provenance & Runtime Specification
- Upstream Checkpoint
- Qwen/Qwen2.5-Coder-0.5B-Instruct
- Upstream License
- Apache-2.0
- Architecture & Active Params
- qwen2 Dense · 0.5B dense
- Certified Context / Max Output
- 32,768 tok / 2,048 tok
- Recommended Hardware Tier
- 4–8 GB CPU-only laptops (~1.34 GiB peak RSS at 32K f16 KV; 2 GB free RAM recommended)
- Weight Modification & Audit Proof
- Derived from Qwen/Qwen2.5-Coder-0.5B-Instruct (Apache-2.0), quantized Q5_K_M GGUF and 4-bit MLX.
Paired Empirical Evidence (Unmerged Comparators)
Artifact-Bound Benchmark Measurements
Micro-Tier Baseline & Memory Profile
Comparator: Single-arm release validation (release_manifest.json)
- BCB First-100
- 7/100
- CPU 32K Peak RSS
- 1.34 GiB
Ultra-small 0.5B micro-model tier
Idle 1.13 GiB at 32K f16 KV
Methodology & Caveat: Intended strictly for syntax helpers, small local completions, and sub-2 GB free-RAM environments.
Live Basis P2 Burn & Effective Customer Rate Across Credit Packs
Certified Sampler Contract (§6.2)
Engine Presets & Cryptographic Checksums
- CRITICAL: MLX build requires strict repeat_penalty = 1.0; LM Studio's default 1.1 degrades/breaks 500m-quill MLX output.
- Windows and Linux deployments support GGUF Q5_K_M only.
llama.cpp (llama-server)
llama-server \
-m verdict-4-500m-quill-Q5_K_M.gguf \
--ctx-size 32768 \
--temp 0.2 \
--repeat-penalty 1.0 \
--port 8080LM Studio (model.yaml)
# Install at ~/.lmstudio/hub/models/verdict/verdict-4-500m-quill/model.yaml
model: verdict/verdict-4-500m-quill
base: verdict/verdict-4-500m-quill-gguf-q5-k-m/verdict-4-500m-quill-Q5_K_M.gguf
config:
load:
fields:
- key: llm.load.contextLength
value: 32768
operation:
fields:
- key: llm.prediction.temperature
value: 0.2
- key: llm.prediction.repeatPenalty
value: 1.0
metadataOverrides:
domain: llm
architectures: [qwen2]
compatibilityTypes: [gguf]
paramsStrings: [0.5B]
minMemoryUsageBytes: 2000000000
contextLengths: [32768]
vision: false
reasoning: false
Verified Local Artifact Checksums
Binary Weights Staged — Gated on Phase 0b Ed25519 SigningSHA-256: da959e399993f511bc00bc72d5da44289bd5d20d2c35c83083df07a14de06656
SHA-256: ee37c7811b9502d7a999818c7f2ac83abb9b30d40fbb40babd3d4f39e231d295
weight_set_sha256: c3caeaf2d711b4a27264c7ba896ecf90c776a26a007f3a93262334f7e416ff6c
Apple Silicon MLX only; requires strict repeat_penalty = 1.0.
Known Tradeoffs & Guardrails
- Local-only micro-model; not hosted on api.verdictmodels.com.
- Breaks on MLX if repeat_penalty exceeds 1.0.