October 2026 · V1 · api.parallelfoundry.com
Parallel Decide
A small, fast model for your app's decisions. It reads text, images and voice, gives every option an honest percentage, explains its answers and calls your tools.
Hi, I was charged twice for my order #4471 this morning. Please refund one of the payments as soon as possible, I need that money for rent this week.
Intent
The user explicitly states, "Please refund one of the payments as soon as possible," which clearly indicates their intent is to request a refund.
Urgent?
The user explicitly states they need the refund "as soon as possible" because they need the money for rent this week, indicating an urgent need.
Real answer from the API1,168 input tokens · $0.0000292
Example answers (real, from the API)
| Input | Question | Answer | Probability |
|---|---|---|---|
| Text: "Hi, I was charged twice for my order #4471 this morning. Please refund one of the payments as soon as possible, I need that money for rent this week." | Intent | refund | 99.6% |
| (same input) | Urgent? | yes | 75.4% |
| Image: A scanned invoice from Peltisdri Group for two office chairs and assembly, EUR 363.00 | Document type | invoice | 99.9% |
| (same input) | Currency | EUR | 100% |
| (same input) | Asks for payment? | yes | 96.1% |
| Recording: phone_call.wav, 8.7 s (MInDS-14 (PolyAI), CC BY 4.0) | What the caller needs | card not working | 100% |
| (same input) | Caller is upset? | no | 55.9% |
| Tool call: "Can you book a table for four people tomorrow at 7:30 pm at Luma Bistro? One of us is vegetarian." (tools: book_table, get_weather, send_email) | Which tool | book_table | 99.95% |
| (same input) | Arguments | restaurant: "Luma Bistro", people: 4, time: "19:30", notes: "One person is vegetarian" | - |
- 74.1%: Zero-shot average on 9 public decision sets
- 25 ms: Per decision on one GPU
- 463: Decisions per second on one GPU
- $0.025: Per million input tokens. Output is free
Capabilities
Decides, explains, and acts.
Every decision model below gives your options a percentage. Parallel Decide is the only one that also explains its answers, proposes options when you have none, and calls your tools.
| Parallel Decide V1 | Jev 1.14 | GPT-6 Luna Decisions | Clef-omni | |
|---|---|---|---|---|
| Reads text | ✓ | ✓ | ✓ | ✓ |
| Reads images | ✓ | ✓ | ✓ | ✓ |
| Hears audio | ✓ | – | – | ✓ |
| Watches video | – | – | – | ✓ |
| A percentage for every option | ✓ | ✓ | ✓ | ✓ |
| Explains its answers | on text | – | – | – |
| Proposes options when you give none | ✓ | – | – | – |
| Calls tools: picks one and writes its arguments | ✓ | – | – | – |
| Finds objects in an image (boxes) | – | ✓ | – | – |
| Open weights | – | – | – | ✓ |
| Context | 10,000 tokens | 262,144 tokens | not published | 64,000 tokens |
| Price per million input tokens | $0.025 | $0.042 | $0.10 | $0.15 |
From each provider's documentation, October 18, 2026: Jev 1.14, GPT-6 Luna Decisions, Clef-omni. Prices are for text; each provider counts images and recordings its own way.
Text
More right answers, out of the box.
Zero-shot: the question and the labels arrive at call time. None of these sets were trained on or used to tune. Ahead of GLiNER2.5-Decide on 8 of 9 sets, and 86.2% to 80.8% on four sets neither model saw in training.
- Parallel Decide
- GLiNER2.5-Decide
- Jev 1.13
- PrismNLI-0.4B
- Laya
| Set | Rows | Parallel Decide | GLiNER2.5-Decide | Jev 1.13 | PrismNLI-0.4B | Laya |
|---|---|---|---|---|---|---|
| Tweet topics | 1,693 | 80.2 | 73.8 | 79.3 | 63.3 | 63.2 |
| Daily dialog | 7,740 | 74.2 | 62.7 | 71.0 | 76.5 | 61.4 |
| Emotion | 2,000 | 57.4 | 55.4 | 58.7 | 72.5 | 58.7 |
| Finance topics | 4,117 | 52.1 | 41.2 | 67.0 | 35.2 | 34.2 |
| SNIPS | 1,400 | 93.1 | 88.2 | – | – | – |
| IMDB | 2,000 | 90.6 | 90.3 | – | – | – |
| SST-2 | 872 | 88.1 | 89.0 | – | – | – |
| AG News | 2,000 | 75.7 | 74.3 | – | – | – |
| 20 Newsgroups | 2,000 | 55.8 | 48.7 | – | – | – |
| Average | 23,822 | 74.1 | 69.3 | – | – | – |
Zero-shot accuracy, %. Best in each row in bold. Parallel Decide and GLiNER2.5-Decide run by us on every row. Jev, PrismNLI and Laya as published by the jev-vs-open-decision-models suite.
Images and recordings
No OCR. No speech-to-text.
Parallel Decide reads the image or the audio itself. Against pipelines that turn it into text first, it is 10 to 18 points ahead on scans, spoken texts, photos and documents, and level on phone calls.
- Parallel Decide, reading the image or audio itself
- Two-step pipeline: OCR, caption or speech-to-text, then GLiNER
| Set | Input | Parallel Decide, % | Two-step pipeline, % | Pipeline | Difference, points |
|---|---|---|---|---|---|
| Real scans (RVL-CDIP) | Image | 53.5 | 35.3 | OCR + GLiNER | +18.2 |
| Spoken texts | Recording | 83.2 | 67.5 | Whisper-small + GLiNER | +15.7 |
| Product photos | Image | 71.7 | 60.9 | Caption + GLiNER | +10.8 |
| Drawn documents | Image | 82.4 | 72.4 | OCR + GLiNER | +10.0 |
| Spoken commands (SLURP) | Recording | 50.5 | 48.5 | Whisper-small + GLiNER | +2.0 |
| Phone calls (MInDS-14) | Recording | 95.1 | 95.5 | Whisper-small + GLiNER | -0.4 |
Speed and cost
25 milliseconds, and output is free.
Every question in a request shares a single pass over the input. You pay only for what you send and what the model reads, never for what it writes.
| Request | Questions | Median at the client |
|---|---|---|
| Decision, short text | 1–3 questions | 24.5 ms |
| Decision, 500 tokens | 3 questions | 29 ms |
| Decision, 2,000 tokens | 3 questions | 89 ms |
| Decision on an image | 2 questions | 162 ms |
| Decision on a recording | 2 questions | 185 ms |
| Written reason | - | 302 ms |
| Tool choice with arguments | - | 400 ms |
Median at the client on one NVIDIA L4, the client in the same region, over 100 requests (25 for the slower kinds). A 10,000-token input takes 0.88 s. From farther away, add your network's round trip.
- $0.025: Per million input tokens, output free. Jev 1.14 costs $0.042, GPT-6 Luna Decisions $0.10, Clef-omni $0.15.
- 463 / s: Short decisions per second on one GPU. A second GPU starts on its own when traffic grows.
- $0: A month. No minimum and no fee: prepaid credit from $5, and refused calls are never charged.
Per million input tokens, output free. Jev 1.14 costs $0.042, GPT-6 Luna Decisions $0.10, Clef-omni $0.15.
Short decisions per second on one GPU. A second GPU starts on its own when traffic grows.
A month. No minimum and no fee: prepaid credit from $5, and refused calls are never charged.
Beyond accuracy
Built to be trusted with a decision.
- 0.007: Calibration error on unseen data. When it says 70%, it is right about 70% of the time.
- 99.0%: The right tool picked on BFCL, its arguments written, and valid JSON on every tool request.
- 0: Texts, images, recordings or answers kept. Nothing you send is stored, logged or used for training.
- 1 pass: For every question in a request: the input is read once, however many questions you ask.
Calibration error on unseen data. When it says 70%, it is right about 70% of the time.
The right tool picked on BFCL, its arguments written, and valid JSON on every tool request.
Texts, images, recordings or answers kept. Nothing you send is stored, logged or used for training.
For every question in a request: the input is read once, however many questions you ask.