Decision models and small local agents: the week AI moved closer to the edge

Liquid AI used 5 October to release d1, a decision model. It answers questions about text or images with probabilities instead of generating tokens. The company's blog says a text decision takes 200 to 300 ms. That is a vendor claim, not an independent benchmark. But the direction is clear: the pitch is latency and price per decision, not raw capability.

Liquid also claims d1 matches or beats GPT-6.1 Sol on four of six tested applications. It says the model costs 19x to 200x less than both that model and Claude Opus 5.5. The company tested d1 on six applications, from filtering support tickets to inspecting circuit boards. It says the model sorts good and defective parts with 85-97% accuracy on a public VisA dataset without being trained for that task. In a coding agent demo, d1 removes 52% of tool-output tokens during context compaction. Per the blog, it keeps every output the task needs.

What a decision model actually does

Liquid describes three question types: Noul (yes/no with a probability), Choice (one label among many) and Score (a position on a scale). One request can carry several questions about the same state, which cuts input tokens.

The vision angle matters more than the text one. Adding a screen raised d1's score in Tetris from 70 to 81 cleared lines. In Wordle it read the board directly from a screenshot, solving 12 of 12 games in 3.8 guesses on average. Those are games, but they show the integration story: no text serialisation of the environment required.

Liquid is not alone in chasing this. Sean Goedecke argued on 5 October that so-called System One models such as Jev are fast general classifiers that any engineering team can prompt. Once a prompt is tuned, he wrote, teams can collect the input/output pairs and train a smaller bespoke classifier from them. His post calls the result "surprisingly easy" to build and much faster. That is a useful counterweight to the vendor framing: the durable win may be distillation, not the general model itself.

Reflection's 501B bet on inference efficiency

Reflection introduced Beam, its first open-weight model, on 5 October. Beam is a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active, according to the company's blog. It was pretrained on 23.8 trillion tokens. The RL run generated more than 100 million rollouts on 10.5K NVIDIA GB300 GPUs over four weeks of training. Reflection says Beam is competitive with larger open models like GLM 5.2 and approaching Qwen 3.8-Max on coding and agentic tasks. It says Beam uses 3 to 4 times less inference compute on advanced reasoning benchmarks.

Read the vendor table carefully. On SWE Bench Pro v2-Hard, Beam scores 77.2 against Kimi K3's 88.2. On Terminal Bench v2.1 it scores 80.1 against Kimi K3's 88.3 and DeepSeek V4.1 Flash's 90.6. Reflection concedes that frontier open models like Kimi K3 remain ahead on raw capability. The weights, technical report and model card are not out yet. The blog says they will be released later this month, and the model is still in final red-teaming. Treat the numbers as claims until then.

Efficiency is the theme that ties the day together. Reflection explicitly frames Beam as "more intelligence per token" for enterprise workhorse use. Liquid frames d1 as a way to replace expensive language model calls when the answer is structured. Both are arguments that the next round of competition is about cost per useful answer.

Small models, small harnesses

Two GitHub releases on 5 October push the same idea down to the laptop. Jeff-Code is a fork of the Pi coding agent by Mario Zechner and contributors. It inserts a 0.8B decision model called Jeff into the agent loop alongside Qwen 3.8-27B. Jeff decides whether it can take the next step itself and whether Qwen needs to think hard on a turn. The repository reports 1,242 paired tasks. Pass rate was 62.4% versus 62.8% for Qwen alone, a paired difference of minus 0.2 points with a 95% interval of minus 2.6 to plus 2.1. Time per task was 0.68x, median 0.70x. About three quarters of Qwen's turns ran with thinking off.

Reika, a coding agent CLI from alexwkleung, takes the opposite starting point. It is designed for 8B to 35B local models at Q2 to Q4 quantisation on a 16 to 32k context window.

Its published findings are refreshingly unflattering. One mid-context rewrite re-processed 8,453 tokens of prompt, 7.9 minutes of prefill. An append in the same session cost 25 tokens and 3.5 seconds. A grep that returned "0 matches" for a call it could not serve sent a model rewording a correct pattern for five rounds. The README says prompt wording is the weakest lever. It also says that on a vague task at Q2 a model converges or spirals about 50/50.

Neither project claims to make small models smarter. Jeff-Code's own caveat is that total time drops less than per-task time. In about 5% of tasks it runs more than 30 minutes longer than Qwen alone. Reika is pre-1.0 and says behaviour can change between releases. Together they are a grounded counterpoint to benchmark tables, because they show where the time and money actually go.

Elsewhere in the stack

The infrastructure releases on 5 October were quieter but relevant to anyone running inference at the edge. Expat 2.9.0 shipped with fixes for CVE-2026-77214, an out-of-bounds read in XML_ParseBuffer, according to the maintainer's blog. It also fixes CVE-2026-102633, an integer overflow affecting 32-bit platforms. The release adds a Properties API and a 64-bit location API. It deprecates the XML_LARGE_SIZE macro that has split ABI compatibility since 2006. The City of Munich's Open Source Sabbatical funded the work, now in its third month of a planned six.

mold 3.0.0 arrived the same day, rewritten from C++ to Rust. The release notes say it is a drop-in replacement for 2.42.1 with the same options, targets and output. They say linking performance is on par. The safety argument is concrete. The C++ version could read out of bounds on corrupted input and segfault. The Rust version bounds-checks those reads and panics at the faulty access instead. The build now uses Cargo and requires Rust 1.95 or later.

On the model side, Interfaze published Interfaze 1.1 Lite, a mixture-of-architectures model. It runs on one 80 GB GPU with no external services. It combines a vision-language reasoning core with specialist models for document reading, speech, layout, segmentation and forecasting. Its model card claims a 95-minute recording transcribes in about 90 seconds. It reports VoxPopuli-Cleaned WER of 3.01, behind Interfaze's own 2.4 but ahead of Gemini-3-Flash's 4.0. NASA and IBM also released fine-tuning code for the NASA-IBM Lunar Foundation Model, with TerraTorch configs for crater detection and ice prospectivity.

None of this adds up to a single product launch. It is a cluster of releases on one day, all pulling in the same direction. Smaller active parameter counts, cheaper per-decision inference, and harnesses that treat context as a budget.

The open question is whether the efficiency claims survive contact with independent testing. For Beam, that starts when the weights land later this month.