← Back to Verdict-4 Model Catalog
Private Beta (Allow-List)Upstream Apache-2.0 Weights + Verdict Sampler/Mode Contract
Verdict 4 20B (Private Beta)
Canonical Cloud SKU: vc/verdict/verdict-4-20b
Serves unmodified upstream openai/gpt-oss-20b Apache-2.0 weights behind Verdict Cloud Gateway's prefill and mode contract on a Modal L4 vLLM container. Currently excluded from the public catalog pending replacement backend qualification.
Lineage, Provenance & Runtime Specification
- Upstream Checkpoint
- openai/gpt-oss-20b
- Upstream License
- Apache-2.0
- Architecture & Active Params
- MXFP4 MoE (21.5B total / ~3.6B active) · 21.5B total / ~3.6B active
- Certified Context / Max Output
- 16,384 tok / 8,192 tok
- Recommended Hardware Tier
- Cloud Private Beta (Modal L4; 125.4s historical cold boot)
- Weight Modification & Audit Proof
- Truthful provenance disclosure (SPEC §2.3 rule 3): unmodified upstream openai/gpt-oss-20b weights governed by Verdict Gateway prefill/mode enforcement.
Paired Empirical Evidence (Unmerged Comparators)
Artifact-Bound Benchmark Measurements
Measured on: Local/LAN LM Studio · Apple Silicon (not Modal L4 Cloud route)
Open-Weights Reasoning Evaluation
Comparator: Warehouse BigCodeBench cohort (N = 3,531)
- BigCodeBench Pass Rate
- 61.1%
N = 3,531 tasks in Verdict warehouse
Methodology & Caveat: Measured on local/LAN LM Studio, not re-measured on the hosted Modal L4 container.
Live Basis P2 Burn & Effective Customer Rate Across Credit Packs
This private-beta SKU is excluded from the unauthenticated public catalog; enrolled allow-list callers receive live rates via authenticated GET /v1/models.
Certified Sampler Contract (§6.2)
Engine Presets & Cryptographic Checksums
- Governed by Verdict Cloud Gateway's existing prefill/mode contract (Qwen enable_thinking chat-template flag does not apply).
- Excluded from public unauthenticated /v1/models; live pricing hydrates only for enrolled allow-list accounts.
Known Tradeoffs & Guardrails
- Private beta (excluded = true in public /v1/models); Modal L4 container exhibits ~125.4s cold starts.
- 16,384-token Cloud context window.