Private AI

Nano GPT logo
NanoGPT
Nano GPT logo
NanoGPT
Nano GPT logo
NanoGPT
Search
 
 
 
 
All conversations

Create

Chat
Studio

Explore

Models
Pricing
AI Apps
Blog

Library

Gallery
Workspace
Usage & billing
API
Suggestions and bugs
Use cases

|

Terms

|

Privacy

|

Refunds
Usage & billing
Teams
Referrals
Presets
Settings
FAQPress & mediaBrand assetsBenchmarksNano RPCEarn crypto
  1. All models overview
  2. Decisions

d3 (Decision Large)

API docs

vLLM Semantic Router’s d3 is the current 27B decision model for complex classification and policy judgments. Runs in an attested TEE (Trusted Execution Environment), with encrypted requests and attestation verified by NanoGPT. It evaluates text against predicate, choice and score questions and returns calibrated probabilities without generating text. Supports a 32,768-token context, up to 64 questions per request, and up to 26 choice options and 10 score levels. Text input only; reasoning, tools and streaming are unsupported. Use the native input and questions-array format through the Decisions API, or supported text questions through System One compatibility. The public ID decision-large is a mutable size alias, not an immutable checkpoint.

Added Oct 22, 2026

Pricing

Auto routing · per 1M tokens
Input
$0.40
Output
Free
Cache read
$0.20
Compare provider prices

Specifications

Context window
32.8K
Image input
No
Questions per request
64

API request

Send your application state and typed questions to POST /api/v1/decisions. Each answer comes back with its probabilities.

{
  "model": "vllm-sr/decision-large",
  "state": "Customer message: I was charged twice for the same order.",
  "questions": {
    "team": {
      "type": "choice",
      "instructions": "Which team should handle this message?",
      "criteria": {
        "billing": "Payments, refunds or invoices",
        "technical": "Bugs or errors in the product"
      }
    }
  }
}
Decisions API docs

Providers

Provider information for this model’s automatic routing. These routes cannot be selected individually.

Choose or restrict providers with the API. Provider-selection prices include the selection markup. Provider selection guide · Pricing API (JSON, USD per 1K tokens)

Auto
Auto
Recommended
$0.40Free$0.20N/AN/A
Tinfoil
Tinfoil
US
$0.40Free$0.20N/AN/A
Privacy:ZDR, prompts not retained

Related decision models

Compare d3 (Decision Large) with similar models from the same provider or model family.

d3 Flash (Decision Medium)

vllm-sr/decision-medium

vLLM Semantic Router’s d3 Flash is the current 9B decision model for triage, routing and policy evaluation. Runs in an attested TEE (Trusted Execution Environment), with encrypted requests and attestation verified by NanoGPT. It evaluates text against predicate, choice and score questions and returns calibrated probabilities without generating text. Supports a 32,768-token context, up to 64 questions per request, and up to 26 choice options and 10 score levels. Text input only; reasoning, tools and streaming are unsupported. Use the native input and questions-array format through the Decisions API, or supported text questions through System One compatibility. The public ID decision-medium is a mutable size alias, not an immutable checkpoint.

d3 Mini (Decision Small)

vllm-sr/decision-small

vLLM Semantic Router’s d3 Mini is the current 4B decision model for fast first-pass classification and high-volume routing. Runs in an attested TEE (Trusted Execution Environment), with encrypted requests and attestation verified by NanoGPT. It evaluates text against predicate, choice and score questions and returns calibrated probabilities without generating text. Supports a 32,768-token context, up to 64 questions per request, and up to 26 choice options and 10 score levels. Text input only; reasoning, tools and streaming are unsupported. Use the native input and questions-array format through the Decisions API, or supported text questions through System One compatibility. The public ID decision-small is a mutable size alias, not an immutable checkpoint.

Bison-1

baselabs/bison-1

Base Labs' Bison-1 is a dense 31B decision model fine-tuned from Gemma 4 31B IT. It returns typed yes/no probabilities, named choices and ordered scores for classification, routing and scoring. Use /v1/decisions or /v1/systemone with text, JSON or up to four inline PNG images, each at most 512 by 512 pixels. Typed questions also work through Chat Completions, Messages and Responses with text input. Supports up to 256 questions, 1–255 choice options, 2–10 score levels and 512 options total; each yes/no question counts as two. NanoGPT uses a conservative 32,768-token admission limit per question. Shared state is billed once, plus each question, including image tokens. Input costs $0.10 per million tokens; answer tokens have no charge. Text generation, reasoning, tools and streaming are unsupported. The base model is Apache-2.0-licensed and the modifications are MIT-licensed.

Groundhog-1

baselabs/groundhog-1

Base Labs' Groundhog-1 is a 26B mixture-of-experts decision model with about 4B active parameters, fine-tuned from Gemma 4 26B A4B IT. It returns typed yes/no probabilities, named choices and ordered scores for classification, routing and scoring. Use /v1/decisions or /v1/systemone with text, JSON or up to four inline PNG images, each at most 512 by 512 pixels. Typed questions also work through Chat Completions, Messages and Responses with text input. Supports up to 256 questions, 1–255 choice options, 2–10 score levels and 512 options total; each yes/no question counts as two. NanoGPT uses a conservative 16,384-token admission limit per question. Shared state is billed once, plus each question, including image tokens. Input costs $0.035 per million tokens; answer tokens have no charge. Text generation, reasoning, tools and streaming are unsupported. The base model is Apache-2.0-licensed and the modifications are MIT-licensed.

Mercury Decide

inception/mercury-decide

Mercury Decide by Inception evaluates shared context and returns binary probabilities, categorical choices, ordinal scores or exact text spans with confidence. Use /v1/decisions with text, JSON or up to four inline PNG images, each at most 512 by 512 pixels. Extraction supports up to 26 fields, uses Unicode character offsets and cannot be combined with images. The context limit is 65,536 tokens, or 32,768 for requests containing extraction, subject to local admission limits. Only input tokens are billed at $0.02 per million; output is free. Chat, streaming and tools are not supported.

Drex Max Finance

nace-ai/drex-max-finance

Nace.AI's Drex Max Finance is a 27B decision model for financial classification, scoring and yes/no judgments. It returns typed Choice, Score and Noul answers with probabilities, without generating text. Use /v1/decisions or /v1/systemone with text, JSON or up to four inline images, each at most 512 by 512 pixels. Supports up to 64 questions, 2–52 choice options and 2–10 score levels. The maximum state context is 262,144 tokens, subject to conservative local admission limits; states over 8,192 tokens are reduced to passages relevant to the questions before processing. Billing uses the input tokens actually read; output is free. Keep meaningful human review for consequential financial decisions.