Skip to main content
← Back to Verdict-4 Model Catalog
Cloud API GA + Staged LocalWeight-Modified Fine-Tune (Gate-W Verified)1.60x Credit Multiplier

Verdict 4 35B ToolTerse v5

Canonical Cloud SKU: vc/verdict/verdict-4-35b-toolterse-v5

Weight-modified derivative of Qwen3.6-35B-A3B trained via rank-32 LoRA SFT (iter 800) across attention and GatedDeltaNet linear-attention projections to eliminate verbose docstring echoing while preserving paired BigCodeBench pass parity.

Lineage, Provenance & Runtime Specification

Upstream Checkpoint
Qwen/Qwen3.6-35B-A3B
Upstream License
Apache-2.0
Architecture & Active Params
qwen3_5_moe (10 Softmax + 30 GatedDeltaNet layers) · 35B total / ~3B active per token
Certified Context / Max Output
32,768 tok / 8,192 tok
Recommended Hardware Tier
>= 32 GB Apple Silicon Unified RAM (MLX) or >= 32 GB CUDA/Vulkan (GGUF); Cloud API for < 32 GB
Weight Modification & Audit Proof
Gate W-GGUF verified: 100/733 tensors modified (self_attn.{q,k,v,o}_proj + linear_attn.{in_proj_qkv,out_proj} across 40 decoder layers), 633/733 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): warm c=1 3,872.5 tok/s prefill / 139.36 tok/s decode; c=4 254.96 tok/s decode window; FlashBoot resume TTFT 8.1-27.5s (cold new worker 43.5-63.7s).

Paired Empirical Evidence (Unmerged Comparators)

Artifact-Bound Benchmark Measurements

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

Token & Wall Efficiency (Comparator 1: Matched Stock GGUF)

Comparator: vs. Matched Stock Qwen3.6-35B-A3B GGUF Q4_K_M (67f8e03f…)

Completion Tokens
-73.15%

Median paired delta, CI95 [-75.22%, -70.82%], p = 3.55e-15 (49 matched-PASS tasks, 3x3 order-balanced)

Wall-Clock Time
-70.91%

Median paired delta, CI95 [-73.06%, -69.39%], p = 3.55e-15

Pass Count (BCB-100)
54 -> 58

49 both pass, b=5 stock-only, c=9 v5-only (63 distinct union passes on first-100); exact McNemar p = 0.4240 — no statistically significant difference detected (not a score win)

Methodology & Caveat: Fused MLX 4-bit sibling measured separately on its own MLX stock control reads -73.29% tokens / -66.78% wall (52 -> 59, gained 12 / lost 5, p = 0.1435). Never merge comparators or lanes.

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

Paired Tool-Use Census (Comparator 2: Predecessor terse-v3)

Comparator: vs. Shipped Predecessor verdict-4-35b-terse-v3 GGUF Q4_K_M

132-Case Tool Census
89/132 -> 116/132

+30.3% over terse-v3 (gained 39 / lost 12, exact McNemar p = 1.98e-4); MLX lane reproduces direction (b=15, c=39, p = 0.0015)

Against Stock Qwen3.6
114/132 -> 116/132

Aggregate parity vs. stock on GGUF (MLX lane: b=14, c=19, p = 0.4869) concealing a stratum split: better abstention on underspecified us_ (p = 0.00024 GGUF) vs. more no_call stalls on determined ds_ (p = 0.00342 GGUF)

Methodology & Caveat: Never claim tool-use superiority over stock Qwen3.6-35B-A3B. Pair with Verdict Proof Mode or Verdict Decision verification when unattended determined tool calls are required.

Measured on: Local MLX 4-bit · LAN LM Studio · Apple M5 Max 128 GB (not Cloud CUDA GGUF)

VDev Direct Agentic Coding Qualification

Comparator: VDev Direct Harness (reasoning=off)

BCB Complete Direct
90.0%

0 reasoning tokens, sub-10s median execution

Warehouse Overall
56.3%

N = 2,717 evaluated tasks across local/LAN warehouse

Methodology & Caveat: Measured on the Apple Silicon MLX 4-bit build, not re-measured on the Cloud RunPod CUDA GGUF Q4_K_M endpoint.

Live Basis P2 Burn & Effective Customer Rate Across Credit Packs

Hydrated live from customer_price_book.v1: raw 1:1 wallet burn is 218.78 cr/MTok input ($0.22 raw) and 6,079.448 cr/MTok output ($6.08 raw).

Credit PackPack Price / CreditsEffective $/CreditCustomer Input $/MTokCustomer Output $/MTok3:1:1 Blended $/MTok
Verdict Starter Pack$5 · 2,000 cr$0.0025$0.55$15.20$8.01
Verdict Plus PackReference 2.0x$10 · 5,000 cr$0.0020$0.44$12.16$6.41
Verdict Ultra Pack$20 · 11,000 cr$0.0018$0.40$11.05$5.83
Verdict Mega Pack$50 · 30,000 cr$0.0017$0.36$10.13$5.34
Verdict Pro PackMost Popular$100 · 62,000 cr$0.0016$0.35$9.81$5.17
Verdict Titan PackBest Rate$200 · 128,000 cr$0.0016$0.34$9.50$5.01

Certified Sampler Contract (§6.2)

Engine Presets & Cryptographic Checksums

  • LM Studio ignores GGUF-embedded sampler metadata and defaults to temperature=0.8, repeatPenalty=1.1; install model.yaml to enforce repeat_penalty=1.0.
  • LM Studio's native MLX chat UI ignores reasoning=off levers; for local MLX use Verdict Local Gateway's offline-render lane or run GGUF via llama-server.
  • Cloud Option B enforces 32,768 tokens per request slot (HTTP 400 exceed_context_size_error above 32K).

llama.cpp (llama-server)

llama-server \
  -m v435btoolterse_v5_iter800_q4_k_m.gguf \
  --ctx-size 32768 \
  --temp 0.2 \
  --repeat-penalty 1.0 \
  --reasoning-budget 0 \
  --jinja \
  --port 8080

LM Studio (model.yaml)

# Install at ~/.lmstudio/hub/models/verdict/verdict-4-35b-toolterse-v5/model.yaml
model: verdict/verdict-4-35b-toolterse-v5
base: verdict/verdict-4-35b-toolterse-v5-gguf-q4-k-m/v435btoolterse_v5_iter800_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: 1.0
metadataOverrides:
  domain: llm
  architectures: [qwen3_5_moe]
  compatibilityTypes: [gguf]
  paramsStrings: [35B-A3B]
  contextLengths: [32768]
  reasoning: false
  vision: false
  trainedForToolUse: true

Verified Local Artifact Checksums

Binary Weights Staged — Gated on Phase 0b Ed25519 Signing
verdict-4-35b-toolterse-v5-gguf-q4-k-mGGUF Q4_K_M · 21,166,757,856 bytes

SHA-256: 7c24da34472201afb21152cd74aca4ed0131c0a2ecb22c6b8e6aae72a32a4530

Post-normfix canonical digest only (v435btoolterse_v5_iter800_q4_k_m.gguf). Reject any pre-normfix digest.

verdict-4-35b-toolterse-v5-mlx-4bitMLX 4-bit · 19,509,023,921 bytes
  • model-00001-of-00004.safetensors (5,343,699,404 B): f750b39e39025b5defebbf4489d939b885db967b2c256921e9cbb4efeb622f2b
  • model-00002-of-00004.safetensors (5,368,472,723 B): 4d9b5f35536938af11ee070ea4c9c10248750e10765c8780d5afe4d9e2a1206a
  • model-00003-of-00004.safetensors (5,368,324,201 B): b91b6f5c7e4c5f81d8690f0e7eff195c7deb6e73c25a274094209462880c0000
  • model-00004-of-00004.safetensors (3,428,527,593 B): 2d6fbb09305a3727308d93a0ffc8478725faa55f668a5fddc18ca2b4abbe2d37

Per-shard digests verified; safetensors-shard-set-v1 weight_set_sha256 pending computation prior to Phase 0b publication.

Known Tradeoffs & Guardrails

  • Loses 5 tasks on BCB-100 that stock Qwen3.6-35B-A3B solves (b=5, c=9); designed for worker tiers backed by test/proof verification.
  • Trades terse-v3's tool over-calling for occasional no_call abstention on fully determined tool requests (ds_ stratum).
  • Hard 32,768-token per-request context limit on Cloud Option B; configure Aider/OpenCode chat-history compaction accordingly.

Call vc/verdict/verdict-4-35b-toolterse-v5 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-4-35b-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)
Verdict 4 35B ToolTerse v5 (vc/verdict/verdict-4-35b-toolterse-v5) · VerdictModels