Skip to main content
TrueSynthesis Inc d/b/a VerdictModels·api.verdictmodels.com

Terse coding & tool-use models built for real execution loops.

First-party Verdict-4 LoRA fine-tunes cut completion tokens by 51.4%–73.15% and wall-clock latency by 43.9%–70.91% against matched stock checkpoints on paired BigCodeBench evaluations—served over an OpenAI-compatible Direct API that shares your existing Verdict Cloud credit wallet with zero double-markup.

Direct API ContractOpenAI v1 Compatible
Base URL
https://api.verdictmodels.com/v1
Auth & Scope
Bearer vrd_live_* + literal "llm:chat"
Wallet & Billing
Shared Verdict Cloud non-expiring credits (1:1 raw burn)
Active Public SKUs
3 live on /v1/models
Prompt Retention
Zero prompt persistence · Zero model training

Live Hydrated Model Catalog

Verdict-4 Family — Cloud SKUs & Local Release Identities

Cloud rates hydrate live from GET https://api.verdictide.com/v1/models?vendor=verdict (customer_price_book.v1). Every empirical metric cites its exact measured quantization, engine, and hardware lane.

Live /v1/models · 19:58:55Z
Cloud GA + Staged Local1.60x tier

Verdict 4 35B ToolTerse v5

vc/verdict/verdict-4-35b-toolterse-v5

Flagship terse MoE coding & tool-use worker (-73.15% tokens vs. matched stock).

Base / License
Apache-2.0
Certified Context
32,768 tok
Provenance
Weight-Modified Fine-Tune (Gate-W Verified)
Customer (Plus ($10 · Ref))$0.44 in / $12.16 out / MTok
1:1 Raw Wallet Burn (Basis P2)218.78 / 6,079.448 cr/MTok

Measured on: Local GGUF Q4_K_M (7c24da34…) · llama.cpp Metal · Apple M5 Max

Completion Tokens: -73.15% (vs. Matched Stock Qwen3.6-35B-A3B GGUF Q4_K_M (67f8e03f…))

Cloud GA + Staged Local1.49x tier

Verdict 4 30B Terse

vc/verdict/verdict-4-30b-terse

Fast 48-layer Softmax MoE coder (-65.26% tokens on shipped GGUF at rp=1.1).

Base / License
Apache-2.0
Certified Context
32,768 tok
Provenance
Weight-Modified Fine-Tune (Gate-W Verified)
Customer (Plus ($10 · Ref))$0.39 in / $11.34 out / MTok
1:1 Raw Wallet Burn (Basis P2)192.862 / 5,667.78 cr/MTok

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

Completion Tokens: -65.26% (vs. Matched Self-Converted Stock Qwen3-Coder-30B GGUF Q4_K_M)

Cloud GA + Staged Local1.59x tier

Verdict 4 4B Quill

vc/verdict/verdict-4-4b-quill

Compact 4B dense model for 12–24 GB laptops and high-concurrency Cloud edits.

Base / License
Apache-2.0
Certified Context
32,768 tok
Provenance
Weight-Modified Fine-Tune (Gate-W Verified)
Customer (Plus ($10 · Ref))$0.23 in / $12.41 out / MTok
1:1 Raw Wallet Burn (Basis P2)115.765 / 6,204.555 cr/MTok

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

Completion Tokens: -51.4% (vs. Matched Stock Qwen3-4B-Instruct-2507 GGUF Q4_K_M)

Local & MVAI ($0/tok)

Verdict 4 1.5B Quill

Ultra-light 1.5B local & MVAI model for < 12 GB CPU-only laptops ($0/tok).

Base / License
Apache-2.0
Certified Context
32,768 tok
Provenance
Weight-Modified Fine-Tune (Local Only)

Local execution only — $0.00/MTok on local CPU/Apple Silicon hardware.

Measured on: Local CPU GGUF Q5_K_M (32K f16 KV) · Apple M4 Max

BCB First-100: 34/100 (Single-arm release validation (release_manifest.json))

Local & MVAI ($0/tok)

Verdict 4 500M Quill

Sub-gigabyte 0.5B local micro-model for 4–8 GB CPU-only machines ($0/tok).

Base / License
Apache-2.0
Certified Context
32,768 tok
Provenance
Weight-Modified Fine-Tune (Local Only)

Local execution only — $0.00/MTok on local CPU/Apple Silicon hardware.

Measured on: Local CPU GGUF Q5_K_M (32K f16 KV) · Apple M4 Max

BCB First-100: 7/100 (Single-arm release validation (release_manifest.json))

Private Beta (Allow-List)

Verdict 4 20B (Private Beta)

vc/verdict/verdict-4-20b

Private-beta MXFP4 MoE reasoning route wrapping upstream Apache-2.0 weights.

Base / License
Apache-2.0
Certified Context
16,384 tok
Provenance
Upstream Apache-2.0 Weights + Verdict Sampler/Mode Contract

Excluded from public unauthenticated catalog; hydrated for enrolled allow-list API keys.

Measured on: Local/LAN LM Studio · Apple Silicon (not Modal L4 Cloud route)

BigCodeBench Pass Rate: 61.1% (Warehouse BigCodeBench cohort (N = 3,531))

Held — Coming Soon

Verdict 4 26B A4B IT (Held / Coming Soon)

vc/verdict/verdict-4-26b-a4b-it

Instruction-specialist Gemma-4 MoE candidate — held pending provider & SHA-256 qualification.

Base / License
Apache-2.0
Certified Context
32,768 tok
Provenance
Upstream Apache-2.0 Weights + Verdict Sampler/Mode Contract

Withheld from commercial billing pending serving-lane qualification (Unlisted / Coming Soon — requests for vc/verdict/verdict-4-26b-a4b-it return 404 until Cloud qualification completes.).

Measured on: Local Mixed 4/8-bit MLX Build (16.4 tok/s) · Apple M4 Max (not Cloud)

Live VDev Assertions: 100% (Internal VDev & Warehouse Evaluation)

Hardware-Fit Advisor

Match your local RAM/VRAM tier + Cloud burst lane

Recommended Local Fit

Verdict 4 35B ToolTerse v5 & Verdict 4 30B Terse

19.71 GiB GGUF Q4_K_M / 18.17 GiB MLX 4-bit (35B v5) · 17.28 GiB GGUF / 16.00 GiB MLX (30B Terse).

Inspect verdict-4-35b-toolterse-v5 preset →

Required Engine Flags

35B ToolTerse v5 uses repeat_penalty=1.0; 30B Terse GGUF uses recertified repeat_penalty=1.1.

Hybrid Cloud Pairing

Mirror identical model behavior between local workstation and Cloud CI via api.verdictmodels.com.

Cryptographic Release Identities & Certified Presets

Local Artifact SHA-256 Ledger & Sampler Downloads

Binary Weights: Staged — Gated on Phase 0b Ed25519 Signing

Certified LM Studio model.yaml, llama-server launch scripts, and canonical SHA-256 digests are live today. Direct public binary weight downloads unlock after Phase 0b detached Ed25519 manifest signing and owner publication sign-off (tracked in VX-2306). Note: ollama is not yet a certified runtime for Verdict-4 GGUFs.

Model / ArtifactFormatExact SizeCanonical SHA-256 / Weight-Set DigestSampler Preset Downloads
Verdict 4 35B ToolTerse v5
verdict-4-35b-toolterse-v5-gguf-q4-k-m
GGUF Q4_K_M19.71 GiB (21,166,757,856 B)7c24da34472201afb21152cd74aca4ed0131c0a2ecb22c6b8e6aae72a32a4530
Verdict 4 35B ToolTerse v5
verdict-4-35b-toolterse-v5-mlx-4bit
MLX 4-bit18.17 GiB (19,509,023,921 B)4 shards verified (f750b39e3902…); weight_set_sha256 pending Phase 0b
Verdict 4 30B Terse
verdict-4-30b-terse-gguf-q4-k-m
GGUF Q4_K_M17.28 GiB (18,556,688,704 B)c9e419cc406ab9c24b798951919567b9fb91aa9cc1ddac265587d40d537c3f4e
Verdict 4 30B Terse
verdict-4-30b-terse-mlx-4bit
MLX 4-bit16.00 GiB (17,174,780,538 B)weight_set: b5cce0d30027942f511e252b6713ef5bc91cbbe8980a2afeda18ba8153a05df5
Verdict 4 4B Quill
verdict-4-4b-quill-gguf-q4-k-m
GGUF Q4_K_M2.33 GiB (2,497,278,880 B)733f3e101ecb1d3cbc14292d9100e2e9c4b869f03674b4aa5ef014d768788abe
Verdict 4 4B Quill
verdict-4-4b-quill-mlx-4bit
MLX 4-bit2.11 GiB (2,263,022,417 B)e6d466b61d25c1e3c26d156221b334226c68c858acd2ed1613f55aebb1558df7
Verdict 4 1.5B Quill
verdict-4-1500m-quill-gguf-q5-k-m
GGUF Q5_K_M1.05 GiB (1,125,050,432 B)604f6903b69bbede5bacd8ac90975db380fd88992e00750c4e98d30785fff0ad
Verdict 4 1.5B Quill
verdict-4-1500m-quill-mlx-4bit
MLX 4-bit0.81 GiB (868,629,082 B)452148f3ad66f7131501827d436475a3e727a8f75e1e1629b9c035c3e92c1dcf
Verdict 4 500M Quill
verdict-4-500m-quill-gguf-q5-k-m
GGUF Q5_K_M0.39 GiB (420,086,144 B)da959e399993f511bc00bc72d5da44289bd5d20d2c35c83083df07a14de06656
Verdict 4 500M Quill
verdict-4-500m-quill-mlx-4bit
MLX 4-bit0.26 GiB (278,065,225 B)ee37c7811b9502d7a999818c7f2ac83abb9b30d40fbb40babd3d4f39e231d295
Verdict 4 26B A4B IT (Held / Coming Soon)
verdict-4-26b-a4b-it-mlx
MLX 4/8-bit MixedPending4 shards verified (…); weight_set_sha256 pending Phase 0b

Drop-In Integration

Connect OpenAI SDKs, Aider, OpenCode, or verdict-code in 30 seconds

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-4-35b-toolterse-v5",
    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)

Live Direct API Playground

Test api.verdictmodels.com in your browser

Key held in tab memory only · Zero storage
Sample prompt

Debits your shared Verdict wallet at 1:1 raw cost. Cold scale-to-zero wakeups exceeding 55s return guidance to call api.verdictmodels.com directly (150s cold-start ceiling).

Illustrative example (not a live run)Enter vrd_live_* key & click Run for live execution
def merge_intervals(intervals: list[tuple[int, int]]) -> list[tuple[int, int]]:
    if not intervals:
        return []
    ordered = sorted(intervals, key=lambda x: x[0])
    merged = [ordered[0]]
    for start, end in ordered[1:]:
        prev_start, prev_end = merged[-1]
        if start <= prev_end:
            merged[-1] = (prev_start, max(prev_end, end))
        else:
            merged.append((start, end))
    return merged