Vega
Nandakishor M's typed-decision model. A frozen Qwen3.5 backbone is read, never trained, and cut off mid-stack; a small trained physics engine settles each choice, score or yes/no question into calibrated probabilities, a conformal set and an abstain flag. 0.8B and 4B.
A personal release by Nandakishor M on his own GitHub, Hugging Face and site, first published on 9 October 2026; not to be confused with vLLM Semantic Router's Decision-2.0-Vega-27B. The backbone, Qwen3.5-0.8B (Qwen3.5-4B for the 4B, under 4b/ in the same repository), is frozen and never trained: it is read at two middle layers (13 and 19 of 24 in the 0.8B), the pass stops there and the LM head never runs. The trained part is an engine with no attention, 14.3M parameters (30.9M for the 4B), plus small gated adapters. It places a particle in a 64-dimensional space with one well per option, runs at most 12 damped steps and reads the answer from where it settles, with a temperature predicted from how it settled. Every answer carries calibrated probabilities, a conformal prediction set, abstain and unbound (out-of-distribution) flags, and which of the two adapters (public_intents, typed_decisions) was used. Context is 73,728 tokens; a state of 4,096 tokens or more is read once and shared by its questions. Images go through the backbone's frozen vision encoder (vegaml's vision extra). An optional head fitted on at least six labelled examples per question is sharper on its own label space but has no calibration or abstain flag. It runs with the maker's vegaml package (Apache-2.0; this revision's card asks for 0.8.0 or later); the backbone downloads separately from Qwen's repository, and this page lists only the data files vegaml reads, no Python files. The maker's own runs against Jev 1.13.0 on identical items, which he calls a selected set of the places Vega leads: phishing screening on 800 emails, 75.4 against 61.9 accuracy (252 against 99 of 400 phishing emails caught); temporal_numeric on JevBench v1.4.2 (a third-party benchmark), 46.7 against 20.0; on the third-party S1MB evaluator, enron-spam 1.000 against 0.920 and ag-news 0.955 against 0.806; calibration error over 5,096 paired decisions, 9.5 against 9.3. He says Jev leads on most of the benchmarks he ran, particularly reranking and multi-step reasoning over long documents. On images, RVL-CDIP-N zero-shot gives 0.793 (Apple Vision OCR plus Jev: 0.896). Across S1MB's 137 benchmarks the engine averages 12.75, and 22.56 with the fitted head. His newer benchmark, run on 11 October 2026 and published with its results and its harness (a notebook in vegaml), covers 43 public tasks plus JevBench (231 items) and DecisionBench (1,075 rows), 3,841 decisions in all with seed 20261011: macro accuracy over the 43 public tasks is 0.556 for the 0.8B and 0.717 for the 4B, and median latency 74 ms and 106 ms per decision on a T4, one question per call.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one probability
- classify — assigns a category from a fixed taxonomy
- route — sends the input to one of several destinations
At a glance
| Parameters | 0.8B |
| Base model | qwen/qwen3.5-0.8b |
| Maker | Nandakishor M |
| Released | 2026-10-09 |
| License | apache-2.0 |
| Reported accuracy | 55.6% |
| Reported latency | 267 ms median per decision on a T4 (fp16) and 280 ms on Apple M-series (fp32) in the maker's Jev comparison (Jev's hosted API: 591 ms); 74 ms median per decision for the 0.8B and 106 ms for the 4B on a T4, one question per call, over the 43 public tasks of his 11 October 2026 benchmark; his harness for the 2,050-decision set logs 22 ms per case for the 0.8B and 28 ms for the 4B with batched extraction (up to 64 sequences and 16,384 tokens a pass), time per case rather than the latency of one request, on hardware he does not record |
Get the weights
pip install systemonemodels
systemone pull nandakishorm/vega
The files are served from the maker's Hugging Face repository, nandakishorm/vega-08b-public-intents at revision 345934b, and verified against the checksums recorded here.
Read more
- Article
- Model card (both sizes)
- Files at revision 345934b
- Code (vegaml)
- vegaml on PyPI
- Demo Space
- Backbone of the 0.8B (Qwen3.5-0.8B)
- Backbone of the 4B (Qwen3.5-4B)
This page was opened by System One for Nandakishor M, who can claim the organisation and take it over at any time.