ENPT

Decision 1.0 Kai 0.6B

Last updated:

Open14–107ms$0/M input

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

  • Maturity5.1/10

    Apache-2.0, ungated, detailed card with architecture, benchmark and training-source attribution; runs in stock Transformers via trust_remote_code with a system_one() call. Serving needs a separately distributed vLLM Semantic Router Decision runtime. No hosted API or SLA.

  • Capability6.6/10

    Noul / Choice / Score in one encoder pass. No calibration fit (temperature 1). Fast: 30 ms median / 107 ms p95 on the official Decision Index for the previous Kai weights. Accuracy is weak: that same run scored 6.52 (#59 of 70); the new weights are only measured by the authors.

  • Adoption4/100

    About 100 downloads and 9 likes on the new weights; the Decision 1.0 family has several entries on the official Decision Index. No integrations or production cases reported.

Vendor claims

Read this first

  1. The only third-party number is low. The official Decision Index scored the previous Kai weights at 6.52 (#59 of 70), just above Laya (6.04) and far from Jev (57.91). The card says the new weights add 7.04 points on the authors’ own suite; they are not on the index yet.
  2. 1,024-token hard limit. A longer question fails the whole request (max_length_exceeded); nothing is truncated.
  3. Probabilities are not calibrated. The card says the calibration slot is unused and temperature stays at 1.
  4. There is a successor. Decision 2.0 Kai-0.6B (2 October) is a different model: a Qwen3-0.6B decoder instead of this encoder, 8,192-token context, and a self-run index score of 16.3. This page stays on the 1.0 weights because its numbers come from the official index.

Decision 1.0 Kai 0.6B is an open decision model from the vLLM Semantic Router project, first published on 21 September 2026 and updated on 1 October (02:19 BRT). It is an encoder, not an LLM: three 22-layer bidirectional paths built on Vela-1.0-Encoder-307M (mmBERT lineage), 571.9M parameters in total. It reads the state and the candidate descriptions together and returns System One answers (Noul, Choice, Score) with probability distributions.

Specs

Attribute Value
Maker vllm-sr (vLLM Semantic Router), formerly llm-semantic-router
Base llm-semantic-router/Vela-1.0-Encoder-307M (mmBERT lineage)
Parameters 571.9M (FP32)
Context 1,024 tokens per question including candidates and state
Question types Noul, Choice, Score
Usage AutoModel.from_pretrained(..., trust_remote_code=True).system_one(...); pipeline("decision"); or a vLLM Semantic Router Decision runtime (distributed separately)
License Apache-2.0 (Gemma-origin tokenizer terms retained)
Launch

Evidence

Source Number
Official Decision Index 0.2.1 (previous Kai weights, 28/09) 6.52 skill, #59 of 70; ECE 0.185; median 30.4 ms / p95 106.8 ms
Authors’ suite (UnverifiedClaim) 53.52 vs Laya English 51.03
Authors’ latency (UnverifiedClaim) 13.6 ms for 1 question; 163 ms for 128 mixed questions (AMD)

The rest of the family on the official index: Decision 1.0 Lux-9B 43.49 (#14), Nox 34.36, Sol 25.32, Eos-0.8B 18.41, Lex 4.54. Lux-9B, the family’s strongest point, has its own page. The whole 2.0 generation is on the Decision 2.0 page.

Fit / anti-fit

Fit: very cheap, very fast local routing and yes/no checks on short inputs, especially multilingual ones (mmBERT base), where you can test accuracy on your own data first.

Anti-fit: long documents; anything that needs calibrated confidence; reasoning-heavy decisions.

Why it is in the catalog

Kai is trained for decisions (its own encoder paths and heads), released openly and answers in the System One format with probabilities. The low Decision Index score is a quality problem, not a classification problem, so it is listed with that number at the top.