Robotics foundation models get the attention, but the decision layer is where the numbers land

On 5 October, Liquid AI released d1, which it calls its first decision model and the first of its kind to handle images alongside text. According to the company's own blog post, d1 answers questions about a situation with a probability per possible answer. It reads text, images or both in a single forward pass without generating tokens. A text decision takes 200 to 300 ms.

The vendor's benchmarks deserve scepticism. Liquid says d1 matches or beats GPT-6.1 Sol on four of six applications it tested, from filtering support tickets to inspecting circuit boards, and costs 19x to 200x less than both that model and Claude Opus 5.5. On the public VisA dataset it sorts good and defective parts at 85 to 97% accuracy, a task, the company says, it was never trained for.

Why does this matter for robotics? Because the embodied AI story of the past year has been dominated by big foundation models, and the harder problem is often the cheap, fast judgement underneath them: is this part defective, is this grasp safe, does this frame need a second look. Liquid's own examples lean that way. It files documents into folders, operates a flight-search site from a one-sentence goal, and compacts a coding agent's tool output, removing 52% of tokens while keeping what the task needs.

Small models doing the deciding

The same pattern shows up in two other releases this week. A GitHub project called Jeff-Code, published 5 October, puts a 0.8B decision model inside the agent loop of Pi, the coding agent by Mario Zechner and contributors. Around every turn of Qwen 3.8-27B, Jeff makes two calls of about 0.2 seconds each: whether it can take the next step itself, and whether Qwen needs to think hard. The project reports a 62.4% pass rate against 62.8% for Qwen alone across 1,242 paired tasks, with time per task at 0.68x and a paired difference of minus 0.2 points.

That is a small sample of small numbers, but the direction is consistent with Liquid's pitch: pay for frontier reasoning only when a cheap model says you need it. Jeff-Code also shows the failure mode. In about 5% of tasks it runs more than 30 minutes longer than Qwen alone, which is why total time only falls to 0.86x.

Reika, another 5 October release, takes the opposite tack. It designs a coding agent CLI around small local models first, in the 8B to 35B range at Q2 to Q4 quantisation on a 16 to 32k context window. Its documentation is unusually candid about the limits. In testing, 14B to 35B models handled multi-file tasks most reliably, 8B to 9B models managed simple, well-scoped ones, and below 7B rarely got far.

The harness's own overhead came up repeatedly. 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. A grep that reported "0 matches" for a call it could not serve sent a model rewording a correct pattern for five rounds.

"Prompt wording is the weakest lever," the Reika README states. On a vague task at Q2 a model converges or spirals about 50/50, and the harness cannot move that rate, only what the failing half costs.

The big open-weight release is still pending

Above all this sits Beam, Reflection's first open-weight model, announced 5 October. It is a sparse mixture-of-experts with 501 billion total parameters and 23 billion active, pretrained on 23.8 trillion tokens, with a reinforcement learning run of more than 100 million rollouts on 10.5K NVIDIA GB300 GPUs over four weeks.

Reflection is careful about what it has and has not shown. Beam is still undergoing final red-teaming and evaluations; the weights, technical report and model card arrive later this month. The company claims Beam is competitive with larger open models like GLM 5.2 and approaching Qwen 3.8-Max on coding and agentic tasks, while acknowledging that Kimi K3 remains ahead on raw capability. Its advantage, the post argues, is inference efficiency, with comparable reasoning scores to GLM-5.2 at 3 to 4 times less compute.

For anyone building robot policies on open weights, that efficiency claim is the one to watch, and it is unverified until the technical report lands.

Primary material, not press releases

Two other items in the dossier are worth reading directly. NASA and IBM published fine-tuning and inference code for a Lunar Foundation Model on GitHub on 5 October, with TerraTorch configs for crater detection, IMP segmentation and ice prospectivity, plus weights on Hugging Face. It is a rare case of a space agency shipping a reusable geospatial backbone with runnable downstream tasks.

Interfaze 1.1 Lite, posted to Hugging Face on 5 October, is a mixture-of-architectures model that runs on one 80 GB GPU with no external services: a reasoning core calling specialist models for OCR, speech, segmentation and forecasting, with structured JSON output. It reports 83.8 on olmOCR-Bench and 73.2 on MMMU-Pro, against 85.7 and 71.1 for Interfaze itself.

Not every number this week supports the small-model thesis. A UC Riverside study published 23 September found that confidence and correctness in Llama-3.1-8B and Gemma-2-9B arise from different internal features, so a model's certainty is not a reliable signal that it is right. Suppressing features tied purely to uncertainty sharply reduced accuracy, according to the university's write-up. Any system that routes work based on a model's own confidence should read that paper first.