Decision Models / REAX
⚠️ Why not a model
Hosted inference runtime over existing open checkpoints; REAX is a serving engine, and the service has no separately published decision-trained weights.
Technical specs
- Base LLMs
- vLLM-SR d3, Blink v0.3, Quyet-1.0-Large, other versioned open checkpoints
- Decision types
- choice, score, noul
- Features
- hosted-api, systemone-compatible, peer-to-peer-tier, eu-hosted-tier, text, image, evidence-page
- License
- Proprietary service; hosted checkpoints retain their own licenses
Links
Decision Models is a hosted inference service tied to REAX, Bittensor subnet 92. The launch post describes an API for open decision models; the current site exposes a typed, Jev-compatible endpoint and separate Peer-to-Peer and EU tiers.
Current service
The public model list names s1-pro (currently vLLM-SR d3), s1-fast and s1-vision (Blink v0.3), and quyet-1-large. These are service aliases over existing checkpoints; versioned model IDs identify which weights answered. The hosted service and REAX engine are not themselves a new catalog model.
The API accepts System One requests and returns typed answers with probabilities. The site currently advertises text and image models, with model-specific limits and pricing. Peer-to-Peer processing may use independent GPU providers outside the EU when enabled for a key; the EU tier keeps requests in the EU. Check the current docs, pricing and status before using the service.
Evidence and disclosure
The founder’s evidence page publishes methods and raw measurements, while stating that the founder runs JevBench and that Decision Models offerings are not listed on that benchmark. In the launch thread, the founder disclosed this role and said REAX/DecisionModels would not influence the evaluation. Keep that disclosure alongside any JevBench claims; the stated independence is not independent verification.
The X launch post advertised up to 4× throughput and a starting price of $0.019 per million input tokens. These are time- and model-specific vendor claims. The evidence page says its speed and quality measurements were self-measured; published leaderboard results refer to underlying open models, not the Decision Models service as a whole.
Launch promotion and current lineup (17 Oct 2026)
publicado (OK do Fabricio 17/10 BRT).
The models page (checked 17 Oct) shows a launch-week promotion: s1-fast-1 and s1-pro-1 free on the REAX Peer-to-Peer tier with code LAUNCH until Sunday 18 Oct, 22:00 UTC (single-question text requests, network-first keys only, rate-limited, may end early). s1-pro now runs vLLM-SR d3 (26B, FP8, EU-hosted); the previous version, Perplexity Decider v1.1, stays available as the pinned id s1-pro-2 until at least 11 November 2026. The four models are labelled Apache-2.0: s1-pro (d3), s1-fast and s1-vision (Blink v0.3 26B-A4B) and quyet-1-large. Promotion terms are the vendor’s and time-limited.
Why this is not a model
The service hosts existing open decision checkpoints and supplies inference infrastructure. No separate Decision Models checkpoint or decision-trained weights are published as the product. Catalog the underlying checkpoints individually when they meet the model criteria; keep this service off the quadrant.