Announced 10 Oct 2026 · Sources checked
What did Nandakishor publish on 10 October?
The dated prose object is the essay “Vega: a decision model that rolls a ball down a hill,” stamped 10 October 2026 on nandakishorm.com. It defines Vega as a decision model: you supply a situation and a question whose answer type is fixed in advance, and you get back a usable value—a choice, a score, or the probability that a statement is true—plus calibrated confidence, a conformal answer set, and an abstain flag.
The software objects match that date window. Hugging Face’s API for nandakishorm/vega-08b-public-intents reports createdAt 2026-10-09T15:16:54Z and lastModified 2026-10-10T19:32:16Z, with cardData license apache-2.0 and base_model Qwen/Qwen3.5-0.8B. GitHub’s API for NandhaKishorM/vegaml reports created_at 2026-10-09T17:45:54Z, pushed_at 2026-10-10T19:33:39Z, and license SPDX Apache-2.0. The LICENSE file we opened is the standard Apache License, Version 2.0.
Typed decision models have been a crowded lane this week. For Microsoft’s Foundry entry, see Microsoft Decision-1. For Nace’s open Drex weights, see Nace Drex v1.5.
How does the frozen reader plus physics engine work?
The essay’s central design claim is a hard split. On one side sits a frozen Qwen3.5 reader (0.8B default, or 4B under a `4b/` tree) that is SHA-256 pinned and never trained. On the other sits a small engine—about 57 MB of trained parameters for the 0.8B path, about 124 MB for the 4B—that does all the learning. The language-modeling head never runs; the forward pass stops after the deepest layer being read.
Situation, question, and every candidate answer are written into one fixed prompt format. One frozen pass produces token vectors. Vega pools five stretches: situation, question, each option, and the final-token state. Those vectors become a world latent, a foresight latent, a typed probe, and an impulse. Two small networks turn that bundle into a starting position and momentum in a 64-dimensional decision space.
Each option’s pooled vector becomes a Gaussian well (center, depth, width) inside a quadratic bowl. A fixed twelve-step damped Hamiltonian integrator rolls the particle; where it settles is the answer, and how it settled sets a predicted temperature used for Boltzmann occupancy probabilities. The essay stresses that the answer set is not frozen at training time: wells are carved from the option text on each request.
How do you run it, and what are the limits?
The GitHub README’s quick start is `pip install vegaml`, then `vegaml.load()` for the 0.8B default or `vegaml.load("4b")` for the larger reader. Inference is aimed at a single T4 or Apple silicon; a Colab notebook is linked from the repository. Context is stated as 73,728 tokens, with shared encoding across multiple questions on one long state and an explicit refuse-rather-than-truncate policy. For the guardrail failure mode that decision APIs share, see Option-Channel attacks on typed decision guardrails.
Limits to keep in the author column: we did not download the engine shards, pin the Qwen SHA, or reproduce phishing, RVL-CDIP-N, or S1MB numbers. The essay itself warns that a short list of wins is not a scoreboard, that unbound flags fire on unfamiliar label spaces, and that a test-time logistic head can beat the engine on labeled tasks while losing calibration and abstain behavior. Production teams should treat the physics metaphor as implementation detail and score the API shape against their own tickets, intents, and image pages.
If you are comparing hosted decision APIs rather than local engines, also read Cloudflare Clef-omni for another non-generative decision path shipping this month.
What should a decision-model team verify first?
Install vegaml in an isolated environment, load the 0.8B checkpoint, and run the README’s billing-ticket example through `decide` so you see choice probabilities, conformal sets, and abstain flags on a known fixture. Separately exercise an image question if your backbone download includes the vision path the essay describes.
Before swapping a production classifier, build a small paired set in your own label space—same options, different correct answers—so you can tell whether the engine is reading the state or leaning on priors. If you fit the library’s test-time head, keep the essay’s warning in view: the head has no abstain flag and can look calibrated while being wrong outside its examples.
This article is an evidence review of the 10 October essay, the Hub card and eval files, and the Apache-2.0 GitHub tree we opened on 10 October 2026. We did not train an adapter or serve Vega in production.
Common questions
Is Vega a chatbot or a generative model?
No. The essay and README describe a typed decision path that returns choices, scores, or statement probabilities. The language-modeling head does not run; no text is generated in the path.
Are the weights open for commercial use?
The GitHub LICENSE and Hub card we opened state Apache License 2.0 for the published package and weights. Confirm the current files and any third-party Qwen terms before shipping a product.
Did Ai Lookout verify the phishing and RVL-CDIP-N numbers?
No. Those rows are author-reported in the 10 October essay. The typed-decisions adapter table is from the Hub eval artifacts. Run your own harness before changing routing or moderation.
What to remember
Vega’s 10 October drop is an Apache-2.0 local decision model that freezes Qwen3.5 and trains a small physics engine to settle answer valleys. Use it as a concrete open alternative in the decision-model wave; keep every benchmark row in the author column until you measure your own workloads.
Sources & further reading
How this story was made
Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.
Our editorial standards
