Fastino ships a 340M decision model and X takes notice

A 340M-parameter model that returns typed decisions instead of text just out-engaged every frontier-lab story in the beat window. Fastino released GLiNER2.5-Decide on September 24 under Apache 2.0, and the announcement post drew roughly 840 likes and 47,700 views, the highest engagement DeAI's morning sweep recorded in the window.

Key facts

• The release is real and inspectable: a 340M encoder on DeBERTa-v3-large, Apache 2.0, on Hugging Face, installable as AutoExtractor.from_pretrained through the gliner2 package.
• Fastino claims the model leads its own Fast Decisions suite at 60.1% average across 17 datasets — a claim we have not independently verified.
• The model card reports a second variant, GLiNER2.5-Decide-1B, at 59.6% on the same suite, with named baselines JevK5 at 57.6%, SemIf (Qwen3.5-4B) at 56.4%, and Laya Router at 46.6%.
• The X announcement drew roughly 840 likes and 47,700 views, the highest engagement in the beat window per the sweep; engagement numbers are point-in-time and not independently verified.
• Fastino reports 38.3 ms end-to-end p50 latency on an NVIDIA V100 and 167.3 ms on a 48-vCPU Intel Xeon for short-document classification — vendor measurements, not independent.

What's driving the conversation

The announcement came from @fastinoAI, with @urchadeDS — Urchade Zaratiana, the model card's first-listed author and first author of the underlying GLiNER2 paper — amplifying it. The conversation that followed was less about the accuracy headline and more about the serving shape: builders read the release as a small, deterministic classifier you can drop in where a generative model is overkill. The engagement numbers came from DeAI's Grok sweep of the window and should be read as momentum, not a verdict.

The substance

What is verifiable: the Hugging Face model card is live, the license is Apache 2.0, the architecture is an encoder (DeBERTa-v3-large) rather than a generative decoder, and the install path (pip install gliner2) works as documented. The card also states what the model is not: it does not reason, explain, or answer open questions. It is a specialist for operational decisions — routing, triage, sentiment, moderation, handoff, agent-completion checks.

What is a claim: every accuracy number. The 60.1% average comes from Fastino's own Fast Decisions suite, an internally generated classification benchmark across 17 datasets with 5,100 test examples. Fastino acknowledges the suite is internal and distinct from JevBench, and that JevK5 is an open reproduction rather than TypeSafe's Jev itself. What would verify it: an independent rerun of the Fast Decisions suite, a third-party leaderboard placement, or per-dataset numbers others can reproduce from the published card. None of those exist yet.

One comparison worth flagging for our own coverage trail: Laya, the open Jev-style decision model we covered last week as a PULSE, sits at 46.6% on Fastino's suite as one of the named baselines. That number is from Fastino's harness, not Laya's authors', and the two teams used different label sets and test conditions.

Why builders are watching

The serving-shape argument is the durable part. An encoder that returns schema-constrained decisions with confidence scores runs on CPU and in air-gapped environments, which puts routing, triage, and tool-call gating in reach for deployments where a chat-completions API is both overkill and a data-transfer event. That is the same argument the Laya release started, from a different team and a different architecture, and it is now happening twice in two weeks — which makes it a pattern rather than a one-off. Whether either model holds up under independent testing is unresolved; how Laya's benchmarks fared under counter-testing is the closest precedent for how this one should be read. The broader open-API landscape that models like this plug into is tracked in our roundup of open-source LLM API providers.