What happened
Fastino announced GLiDE on November 5, 2026, as a family of five decision models: 0.4B, 0.8B, 2B, 9B, and 27B. The company describes decision tasks as cases where the application already knows the possible answers: which queue should receive a ticket, whether a request is safe, whether to call a tool, or how to rate a response. GLiDE reads the context, question, and candidate options together, then returns a probability for each option in one forward pass, without generating tokens. All five sizes use the same System One schema, so an application can switch model size without changing its questions or answer parsing.
The release table reports expected Full scores on Decision Index 0.4.1, with 80% intervals, alongside measured public-suite scores, median latency, and bf16 memory. GLiDE 27B is listed at an expected Full score of 73.3 (71.4–75.2), a public-suite score of 72.5, 112 ms median latency, and 52 GB memory. For 9B, the corresponding figures are 61.8 (57.4–65.5), 68.8, 39 ms, and 16 GB. The 2B is listed at 44.5 (40.6–47.7), 60.6, 18 ms, and 3.8 GB; 0.8B at 36.2 (32.3–39.4), 54.3, 19 ms, and 1.5 GB; and 0.4B at 30.3 (25.2–33.9), 51.1, 7 ms, and 0.8 GB. Latencies were measured on a single NVIDIA RTX PRO 6000; public-suite scores used the official evaluation harness.
Fastino says GLiDE 27B and 9B use recurrent depth: selected middle layers run twice with shared weights, adding computation without adding parameters or producing text. The company says the feature raises the 27B public-suite score from 72.1 to 72.5 and its expected Full score from 72.1 to 73.3, while median latency rises from 83 ms to 112 ms. At 9B, Fastino reports a smaller but statistically reliable gain on held-out evaluations. At 2B, it says a roughly one-point gain came with latency increasing from 18 ms to 28 ms, so that version ships without recurrent depth; below 1B, it found no reliable gain.
The weights are available on Hugging Face. Fastino also makes GLiDE 27B available through its API. The supplied release gives no prices or API limits. Fastino’s comparisons use official peer scores as of November 4, 2026, but GLiDE’s Full scores are explicitly estimates; the company predicts that four sizes will beat similarly sized peers and that 27B will lead overall.
What the archive adds
[Liquid AI’s LFM3 release](/posts/liquid-ai-launches-lfm3-an-open-weight-2b-d00d7d/) also covered an open-weight family that included decision models, but it was a different company and release. [Vectorize’s Hindsight Models](/posts/vectorize-releases-hindsight-models-open-weight-decision-rerank-d1b303/) covered open-weight models for agent memory, not this decision-model family. GLiDE is a new event, not a restatement of either release.
Trend signal
This supports agent economics moves to completed work only weakly: GLiDE targets decision steps that can sit inside a larger workflow, but the release does not report cost per verified completed task. Its more concrete claim is narrower: predefined-option decisions can return calibrated probabilities with fixed-shape output and latency that Fastino measured across model sizes. Builders still need task-level evidence before concluding that this reduces workflow cost.
Why now
Fastino frames the release as an effort to move useful computation for difficult decisions inside the model, instead of spending it on generated reasoning tokens. It also points to recurrent depth as a recently prominent research direction. The source offers that rationale, but does not establish that a market shift or external event caused the timing.
Why it matters
For teams building classifiers, ticket routers, safety gates, or tool-selection systems, a probability distribution over a known set of options is easier to threshold and inspect than free-form text that must be parsed. A shared schema across five sizes could also make it practical to test whether a smaller model meets a particular workload’s accuracy and latency needs without rewriting application logic. The reported 7 ms latency and sub-1 GB bf16 footprint for 0.4B make local experimentation plausible for some setups, while the 27B results illustrate the quality-versus-compute trade-off.
These are not deployment guarantees. The useful comparison is not simply “decision model versus chat model,” but end-to-end performance on the team’s own inputs: error costs, calibration, throughput under load, hardware, and fallback behavior all matter. Treat GLiDE as a candidate to evaluate in a narrow decision stage, not as a reason to replace a general model across a workflow.
Limits and uncertainty
The primary source is Fastino’s own release and evaluation, not an independent replication. Its expected Decision Index Full scores are forecasts with intervals; they are not the official leaderboard results reported for peer models. The release gives measured public-suite figures, but the excerpt does not provide enough detail to assess dataset composition, confidence calibration across deployment domains, or performance under distribution shift. The latency figures are medians from one NVIDIA RTX PRO 6000 and do not establish production throughput, tail latency, or performance on other hardware. No price, API quota, licence terms, or deployment support details are supplied here. The source also does not establish that probability outputs will be well calibrated for a particular team’s data or that one-pass inference will outperform an existing pipeline end to end.
Reader takeaway
GLiDE is worth a controlled test when an application must choose among known options and generated prose is overhead. Its shared interface, probability outputs, and reported latencies give teams concrete evaluation dimensions; the headline ranking remains Fastino’s estimate. Measure accuracy, calibration, and full workflow latency on representative decisions before adopting it.
Open question
How do GLiDE’s calibration and task-level costs compare with a prompted general-purpose model on independent production workloads?