Skip to main content
← Back to Verdict-4 Model Catalog
Cloud API GA + Staged LocalWeight-Modified Fine-Tune (Gate-W Verified)1.59x Credit Multiplier

Verdict 4 4B Quill

Canonical Cloud SKU: vc/verdict/verdict-4-4b-quill

Fused 200-iteration LoRA SFT on Qwen3-4B-Instruct-2507 trained on 469 execution-verified final-code examples. Certified for 32K context on CPU/Metal with Flash Attention enabled.

Lineage, Provenance & Runtime Specification

Upstream Checkpoint
Qwen/Qwen3-4B-Instruct-2507
Upstream License
Apache-2.0
Architecture & Active Params
qwen3 Dense (36 layers, 8 KV heads) · 4B dense
Certified Context / Max Output
32,768 tok / 8,192 tok
Recommended Hardware Tier
12–24 GB Unified/System RAM (with flashAttention: true) or 8–16 GB VRAM
Weight Modification & Audit Proof
Derived from Qwen/Qwen3-4B-Instruct-2507 (rev cdbee75f17c01a7cc42f958dc650907174af0554, Apache-2.0) via seed-0 fused LoRA SFT.

Measured on RunPod Serverless CA-MTL-1 (48 GB A40, llama-server GGUF Q4_K_M, 8x32K slots): warm c=1 7,318.5 tok/s prefill / 136.55 tok/s decode; c=8 288.49 tok/s decode window (381.68 tok/s post-max_ttft); FlashBoot resume TTFT 8.0-21.5s.

Paired Empirical Evidence (Unmerged Comparators)

Artifact-Bound Benchmark Measurements

Measured on: Local CPU llama-cpp-python (n_gpu_layers=0, n_ctx=32768, temp=0.2, FA off on short prompts) · Apple M4 Max

CPU Serving Efficiency (BigCodeBench Tasks 0–199)

Comparator: vs. Matched Stock Qwen3-4B-Instruct-2507 GGUF Q4_K_M

Completion Tokens
-51.4%

-47.8% summary catalog token delta across evaluation cohorts

Wall-Clock Time
-43.9%

Serialized paired CPU evaluation

Pass Count (BCB 0–199)
88/200 vs. 87/200

Paired flips 21:20, exact McNemar p = 1.0 — no statistically significant pass-rate difference detected (no formal non-inferiority margin claimed)

Methodology & Caveat: 30K-token memory recertification (VX-2340) proved Flash Attention ON reduces CPU RSS from 18.64 GiB to 9.42 GiB; BCB quality with Flash Attention ON was not separately re-run.

Live Basis P2 Burn & Effective Customer Rate Across Credit Packs

Hydrated live from customer_price_book.v1: raw 1:1 wallet burn is 115.765 cr/MTok input ($0.12 raw) and 6,204.555 cr/MTok output ($6.20 raw).

Credit PackPack Price / CreditsEffective $/CreditCustomer Input $/MTokCustomer Output $/MTok3:1:1 Blended $/MTok
Verdict Starter Pack$5 · 2,000 cr$0.0025$0.29$15.51$7.97
Verdict Plus PackReference 2.0x$10 · 5,000 cr$0.0020$0.23$12.41$6.38
Verdict Ultra Pack$20 · 11,000 cr$0.0018$0.21$11.28$5.80
Verdict Mega Pack$50 · 30,000 cr$0.0017$0.19$10.34$5.32
Verdict Pro PackMost Popular$100 · 62,000 cr$0.0016$0.19$10.01$5.14
Verdict Titan PackBest Rate$200 · 128,000 cr$0.0016$0.18$9.69$4.98

Certified Sampler Contract (§6.2)

Engine Presets & Cryptographic Checksums

  • MANDATORY (VX-2340): Enable Flash Attention (llm.load.llama.flashAttention: true / --flash-attn). At ~30K prompt tokens on CPU f16 KV, peak RSS is 18.64 GiB with Flash Attention OFF vs. 9.42 GiB with Flash Attention ON (7.1 GiB on Metal FA).
  • Install model.yaml under ~/.lmstudio/hub/models/verdict/verdict-4-4b-quill/model.yaml — never nest model.yaml inside ~/Models/verdict/.

llama.cpp (llama-server)

llama-server \
  -m verdict-4-4b-quill-Q4_K_M.gguf \
  --ctx-size 32768 \
  --flash-attn \
  --temp 0.2 \
  --repeat-penalty 1.0 \
  --reasoning-budget 0 \
  --jinja \
  --port 8080

LM Studio (model.yaml)

# Serving profile cpu-32k-f16-v1 (VX-2340: flashAttention ON required for 12 GB free RAM at 32K context)
# Install at ~/.lmstudio/hub/models/verdict/verdict-4-4b-quill/model.yaml
model: verdict/verdict-4-4b-quill
base: verdict/verdict-4-4b-quill/verdict-4-4b-quill-Q4_K_M.gguf
config:
  load:
    fields:
      - key: llm.load.contextLength
        value: 32768
      - key: llm.load.llama.flashAttention
        value: true
  operation:
    fields:
      - key: llm.prediction.temperature
        value: 0.2
      - key: llm.prediction.repeatPenalty
        value: 1.0
metadataOverrides:
  domain: llm
  architectures: [qwen3]
  compatibilityTypes: [gguf]
  paramsStrings: [4B]
  minMemoryUsageBytes: 12000000000
  contextLengths: [32768]
  vision: false
  reasoning: false
  trainedForToolUse: true

Verified Local Artifact Checksums

Binary Weights Staged — Gated on Phase 0b Ed25519 Signing
verdict-4-4b-quill-gguf-q4-k-mGGUF Q4_K_M · 2,497,278,880 bytes

SHA-256: 733f3e101ecb1d3cbc14292d9100e2e9c4b869f03674b4aa5ef014d768788abe

verdict-4-4b-quill-mlx-4bitMLX 4-bit · 2,263,022,417 bytes

SHA-256: e6d466b61d25c1e3c26d156221b334226c68c858acd2ed1613f55aebb1558df7

weight_set_sha256: de9f8ca6445f9072d9071d9b6cb360d9278a6237b972f7837264a832d837c9a9

Known Tradeoffs & Guardrails

  • 4B dense capacity is suited for bounded edits, boilerplate elimination, and laptop offline fallback rather than deep multi-file architecture.
  • Requires Flash Attention ON to stay within 12 GB free RAM at full 32K context.

Call vc/verdict/verdict-4-4b-quill via Direct API

Sample promptUses standard openai SDK with Idempotency-Key, 160s cold-start safe timeout, and streaming usage.
import os, uuid
from openai import OpenAI

client = OpenAI(
    base_url="https://api.verdictmodels.com/v1",
    api_key=os.environ["VERDICT_API_KEY"],  # vrd_live_* with literal "llm:chat" scope
    timeout=160.0,
)

stream = client.chat.completions.create(
    model="vc/verdict/verdict-4-4b-quill",
    temperature=0.2,
    stream=True,
    stream_options={"include_usage": True},
    extra_headers={"Idempotency-Key": str(uuid.uuid4())},
    messages=[
        {"role": "user", "content": "Write a concise Python binary_search(nums, target) function."}
    ],
)
for chunk in stream:
    if chunk.choices and chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="", flush=True)
Verdict 4 4B Quill (vc/verdict/verdict-4-4b-quill) · VerdictModels