Skip to main content
← Back to Verdict-4 Model Catalog
Cloud API GA + Staged LocalWeight-Modified Fine-Tune (Gate-W Verified)1.49x Credit Multiplier

Verdict 4 30B Terse

Canonical Cloud SKU: vc/verdict/verdict-4-30b-terse

Fused rank-32 attention-only LoRA SFT (iter 200) on Qwen3-Coder-30B-A3B-Instruct trained on 753 execution-verified shortest-passing BigCodeBench solutions. Produces concise direct code without thinking traces.

Lineage, Provenance & Runtime Specification

Upstream Checkpoint
Qwen/Qwen3-Coder-30B-A3B-Instruct
Upstream License
Apache-2.0
Architecture & Active Params
qwen3_moe (48 Softmax layers, 128 experts, 8 active) · 30B total / ~3B active per token
Certified Context / Max Output
32,768 tok / 8,192 tok
Recommended Hardware Tier
>= 32 GB Unified/System RAM or >= 32 GB CUDA; 24 GB discrete GPUs require reduced single-slot context
Weight Modification & Audit Proof
Gate W-GGUF verified on shipped c9e419cc…: 192/579 attention tensors modified across 48 blocks, all 387 MoE expert/norm/head tensors byte-identical to matched stock GGUF.

Measured on RunPod Serverless CA-MTL-1 (48 GB A40, llama-server GGUF Q4_K_M, 4x32K slots, --repeat-penalty 1.1): warm c=1 4,392.9 tok/s prefill / 149.48 tok/s decode; c=4 244.64 tok/s post-prefill decode; FlashBoot resume TTFT 7.9-19.6s.

Paired Empirical Evidence (Unmerged Comparators)

Artifact-Bound Benchmark Measurements

Measured on: Local GGUF Q4_K_M (c9e419cc…) · llama.cpp Metal · Apple M5 Max & M4 Max

Recertified Shipped GGUF Efficiency (BCB-100 at repeat_penalty = 1.1)

Comparator: vs. Matched Self-Converted Stock Qwen3-Coder-30B GGUF Q4_K_M

Completion Tokens
-65.26%

Paired matched-PASS stratum, p = 2.53e-10 (M4 Max replay: -64.77%, p = 1.17e-10)

Wall-Clock Time
-59.72%

Paired matched-PASS stratum, p = 2.53e-10

Pass Count (BCB-100)
60/100 vs. 59/100

Exact McNemar p = 1.0 (M4 Max replay: 62/100 vs. 61/100, gained 7 / lost 6, p = 1.0) — no statistically significant score difference detected

Methodology & Caveat: Efficiency model, not an accuracy lift. Earlier rp=1.0 evaluation scored 62/100 vs. 59/100 (gained 7 / lost 4, p = 0.5488).

Measured on: Local MLX 4-bit · Apple M5 Max

Held-Out BigCodeBench-Hard-148 & Off-Distribution CRUXEval

Comparator: vs. Matched Stock MLX 4-bit

BCB-Hard-148 Tokens
-48.8%

Wall -46.5%; pooled 237-task net score delta -2 (exact McNemar p = 0.8804)

CRUXEval-400 Delta
0.00%

77.1% byte-identical length on literal-answer prompts with no docstring to echo

Methodology & Caveat: Prompts without long docstrings or boilerplate expectations yield ~0% token reduction by design.

Live Basis P2 Burn & Effective Customer Rate Across Credit Packs

Hydrated live from customer_price_book.v1: raw 1:1 wallet burn is 192.862 cr/MTok input ($0.19 raw) and 5,667.78 cr/MTok output ($5.67 raw).

Credit PackPack Price / CreditsEffective $/CreditCustomer Input $/MTokCustomer Output $/MTok3:1:1 Blended $/MTok
Verdict Starter Pack$5 · 2,000 cr$0.0025$0.48$14.17$7.45
Verdict Plus PackReference 2.0x$10 · 5,000 cr$0.0020$0.39$11.34$5.96
Verdict Ultra Pack$20 · 11,000 cr$0.0018$0.35$10.31$5.42
Verdict Mega Pack$50 · 30,000 cr$0.0017$0.32$9.45$4.96
Verdict Pro PackMost Popular$100 · 62,000 cr$0.0016$0.31$9.14$4.80
Verdict Titan PackBest Rate$200 · 128,000 cr$0.0016$0.30$8.86$4.65

Certified Sampler Contract (§6.2)

Engine Presets & Cryptographic Checksums

  • Shipped GGUF c9e419cc… was recertified on BCB-100 at repeat_penalty=1.1 (penalty_last_n=64) in VX-2358 / VX-2355; pin repeat_penalty=1.1 locally.
  • Four 32K KV slots require ~31.5 GB VRAM; on 24 GB discrete GPUs run --parallel 1 with a smaller tested --ctx-size.

llama.cpp (llama-server)

llama-server \
  -m verdict-4-30b-terse-Q4_K_M.gguf \
  --ctx-size 32768 \
  --temp 0.2 \
  --repeat-penalty 1.1 \
  --repeat-last-n 64 \
  --reasoning-budget 0 \
  --jinja \
  --port 8080

LM Studio (model.yaml)

# Install at ~/.lmstudio/hub/models/verdict/verdict-4-30b-terse/model.yaml
model: verdict/verdict-4-30b-terse
base: verdict/verdict-4-30b-terse-gguf-q4-k-m/verdict-4-30b-terse-Q4_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: 1.1
metadataOverrides:
  domain: llm
  architectures: [qwen3_moe]
  compatibilityTypes: [gguf]
  paramsStrings: [30B-A3B]
  contextLengths: [32768]
  reasoning: false
  vision: false
  trainedForToolUse: true

Verified Local Artifact Checksums

Binary Weights Staged — Gated on Phase 0b Ed25519 Signing
verdict-4-30b-terse-gguf-q4-k-mGGUF Q4_K_M · 18,556,688,704 bytes

SHA-256: c9e419cc406ab9c24b798951919567b9fb91aa9cc1ddac265587d40d537c3f4e

Re-issued verified identity (VX-2355, +64 bytes GGUF KV penalty_repeat=1.1 metadata over superseded 2026-08-22 digest d0fff10e…; 579/579 tensors byte-identical).

verdict-4-30b-terse-mlx-4bitMLX 4-bit · 17,174,780,538 bytes

weight_set_sha256: b5cce0d30027942f511e252b6713ef5bc91cbbe8980a2afeda18ba8153a05df5

  • model-00001-of-00004.safetensors (5,319,507,334 B): 96476d168ba2d0cdad1b8e2b11ff02d7d4a56f0e45fb0d32c6562c2211b11477
  • model-00002-of-00004.safetensors (5,364,678,700 B): 98277db12d860e9ec50dc12c2dfaedcb59beb629ece5ef8abfbe918895feb87c
  • model-00003-of-00004.safetensors (5,274,790,267 B): 6c01ecc09e415fb5efa7fcc80cf8b5e94699bb97099301187a6a84b5f237358b
  • model-00004-of-00004.safetensors (1,215,804,237 B): 22588bac8df9d921c68adbb25a4678c95864da99032f8b8a6fd810d3c0512409

safetensors-shard-set-v1 weight_set_sha256 verified read-only in VX-2355.

Known Tradeoffs & Guardrails

  • Trained on shortest-passing solutions: cuts comment bytes (-99.4%), plot-metadata calls like set_title/legend (-53.5%), and try blocks (-36.6%).
  • 78% of paired task regressions are omitted specification details unguarded by unit tests; use alongside automated test or proof gates.

Call vc/verdict/verdict-4-30b-terse 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-4-30b-terse",
    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 30B Terse (vc/verdict/verdict-4-30b-terse) · VerdictModels