'Decision models' promise faster, cheaper AI, but they are not the security tool yet

Liquid AI published d1 on 5 October and calls it its first decision model. It now handles text and images. The company says d1 matches or beats GPT-6.1 Sol on four of six real applications it tested, from filtering support tickets to inspecting circuit boards. It costs 19x to 200x less than that model and Claude Opus 5.5. According to the blog post, it answers significantly faster on every task.

The pitch is narrow by design. Liquid explains decision models this way: they take unstructured data and one or more questions, read them in one forward pass and return probabilities without generating any tokens. A text decision takes 200 to 300 ms. The model answers yes/no questions, picks one label among many, or scores a position on a scale.

That is where security teams might plausibly look. The company's demo list includes a smart filter that turns d1 into a function inside a SQL query, answering yes or no on 150 support tickets. It also lists a context compaction task. There d1 reads each tool output in a coding agent's session and keeps, trims or drops it, removing 52% of the tokens while keeping every output the task needs. Neither is a defence product. Both are the sort of plumbing that sits under one.

A crowded week for small, fast models

Liquid's release landed in a rush of announcements about models that do less, faster. Reflection published Beam on 5 October, a sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active. It is built for coding, reasoning and agentic workloads. The company says it trained on 23.8 trillion tokens. It also ran a high-compute reinforcement learning phase generating over 100 million rollouts on 10.5K NVIDIA GB300 GPUs over four weeks. Beam is still undergoing final red-teaming and evaluations. Reflection says it will release weights, a technical report and a model card later this month.

The efficiency claim is the interesting one. Reflection writes that Beam reaches scores comparable to GLM-5.2 on advanced reasoning benchmarks while using 3 to 4 times less inference compute. It says the gap widens against models in the 2T+ parameter family such as Qwen 3.8-Max. It also concedes the limits. On SWE Bench Pro v2-Hard Beam scores 77.2 against Kimi K3's 88.2, and on Terminal Bench v2.1 it gets 80.1 against 88.3. The post says frontier open models remain ahead on raw capability, and Beam's advantage is efficiency at inference time.

Efficiency arguments matter to anyone running classifiers over large volumes of logs, alerts or tickets. They matter less if the model is wrong.

The confidence problem nobody has solved

A UC Riverside study published on 23 September found that confidence and correctness in large language models can arise from different internal features. That challenges the assumption that a model's confidence signals accuracy. The researchers studied Meta's Llama-3.1-8B and Google's Gemma-2-9B. They sorted answers into four groups by correctness and confidence, then used sparse autoencoders to see which features activated. The university's write-up says suppressing features tied purely to uncertainty sharply reduced accuracy, suggesting those features help produce good answers.

"The main assumption in the field is that when the model is confident, it is likely to be correct, and when the model is unsure, it is more likely to be incorrect," said Het Patel, a UCR computer science doctoral student and lead author of the study. "But we often see the counterexamples that are well documented."

For a decision model returning a probability, that finding cuts both ways. A probability is at least explicit, unlike a sentence delivered with unearned certainty. But the researchers' point is that the number may not track what practitioners assume it tracks.

Two open-source coding agent projects published on 5 October make a related argument about where small models fit. Jeff-Code, a fork of Pi by Mario Zechner and contributors, puts a 0.8B decision model inside the agent loop. It decides whether the agent can take the next step itself and whether Qwen 3.8-27B needs to think hard on a turn. Its README reports coding tasks 47% faster on average at the same pass rate. It reports a paired difference of minus 0.2 points across 1,242 paired tasks, and a 95% interval of minus 2.6 to plus 2.1. Thinking off on every turn was faster still but cost 7.6 points.

Reika, a coding agent CLI designed around small local models first, published its own findings file. Among them: one mid-context rewrite re-processed 8,453 tokens of prompt, 7.9 minutes of prefill, where an append in the same session cost 25 tokens and 3.5 seconds. The project says prompt wording is the weakest lever. On a vague task at Q2, it adds, a model converges or spirals about 50/50. The harness cannot move that rate, only the cost of the failing half.

What none of this is

None of these releases is a cyber defence model, and it is worth being blunt about that. The most recent security-adjacent item is an OpenAI announcement on 3 November about partnerships with security firms, which this piece cannot substantiate beyond its headline. Liquid's d1 is a classifier. Reflection's Beam is a general open-weight model awaiting red-teaming. Jeff-Code and Reika are agent harnesses.

The connection is cost and latency, and it is unglamorous. A decision model can answer a structured question in 200 to 300 ms at a fraction of the price of a frontier call. If so, the economics of running a model over every alert change. Whether the answers are good enough to act on is a separate question. The UCR work suggests the confidence number attached to them deserves scrutiny rather than trust.

Liquid says d1 was never trained for visual inspection and still sorts good and defective parts with 85 to 97% accuracy on the public VisA dataset, understanding the task from a short description. That is an impressive generalisation claim from a vendor blog. It is also exactly the kind of claim that needs an independent test before anyone points it at a production line, or a security operations centre.