Drex: an open-weight 8B model tops the decision-model index
Nace.AI's diffusion model runs on your own GPU and leads the Decision Index, as Upstage ships a 512K-context decision endpoint that bills output at zero.
Copy markdown
Drex: a decision model you download, not rent
Nace.AI's Drex is an 8B diffusion LM with a pointer head (built on NVIDIA's Efficient-DLM-8B) that returns a probability for each typed option — choice, yes/no, or score — in a single forward pass. It leads the six-model Decision Index 0.2 at 52.31, topping retrieval/classification (54.25) and language understanding (57.77). Weights are on Hugging Face (32K context, ~16GB in bf16), so the routing layer runs on your own GPU — but the CC BY-NC 4.0 license rules out paid products (the code is MIT).
Solar Decide Flash: 512K context, output billed at $0
Also new this week, Upstage's Solar Decide Flash (built on Solar Mini 4) answers typed questions in one pass — output tokens are free, input is $0.05/M at a 50% launch discount. A 512K context and ~0.84s median latency make it a cheap router/classifier for long docs; it's served through the System One (/v1/systemone) decisions API, not chat completions.
Build this weekend: a model-agnostic gating layer
Drex, Solar Decide, Cloudflare's Clef and OpenAI's Decisions API now share a similar typed-decision surface, so you can prototype an agent router or policy check against self-hosted Drex and swap to a hosted endpoint for scale without rewriting calls. These return calibrated probabilities instead of prose — exactly what a “should this agent act?” gate needs.