What is Decision-2.0-Nox-4B (vLLM Semantic Router)?
Decision-2.0-Nox-4B is the 4B text model of Decision 2.0, the decision models of the vLLM Semantic Router project, published on Hugging Face on 29 September 2026. Like Jev, it writes no text: you send an input (text or JSON) and named questions, and it answers choice, yes/no and score questions together in one forward pass with a probability for every option.
It is a fine-tune of Qwen3.5-4B-Base with 4.21 billion parameters, a 16,384-token context and Apache 2.0 weights. It reads text only. The card reports a median of 12.9 ms per single-question request on one GPU.
Ten days later the same team published Decision 3.0, whose 4.5B d3-mini is also built on Qwen3.5-4B. vLLM Semantic Router's larger d3 and d3-turbo are reviewed separately.
Is Decision-2.0-Nox-4B (vLLM Semantic Router) worth it?
Decision-2.0-Nox-4B is a sound small model that arrived just before its replacement. On the community Jev Decision Index 0.4.1, data dated 29 October, it scores 40.75 and ranks 134th of 247 entries. That beats the two most comparable 4B models, JPT-4B at 39.56 and Decider 4B at 36.3, which matches the card's claim that it leads its size class. But the board also lists vLLM Semantic Router's own d3-mini, a 4.5B model on the same Qwen3.5-4B base, at 49.97, and Perplexity's pplx-decider-flash-v1.2 at 54.29.
The card's figures are self-reported but carefully framed: 63.6 on JevArena against 61.9 for Decider 4B, described as "statistically level", 52.3 on 15 human-labelled tasks where Decider 4B is ahead at 55.5, and 43.8 on the Jev Decision Index from the team's own run of the 0.2.1 kit. The team says it audited its training data row by row against the index test items, which is more than most decision-model cards claim.
Calibration is its best quality. The community index measures an expected calibration error of 0.030, lower than Jev 1.13 (0.078), d3-mini (0.056) and Decider 4B (0.084). Its probabilities track how often it is right, so a confidence threshold set on your own data should hold. Speed is less impressive than the card suggests: the index recorded a 33.5 ms median on an RTX PRO 6000, against 12.6 ms for Decider 4B.
It reads text only and has no image path, unlike d3-mini or pplx-decider-flash. For a new text-only deployment under a permissive licence, d3-mini scores higher at a similar size; Nox remains a reasonable choice where calibrated probabilities matter more than the last nine points of accuracy.
Where it earns its place
A platform engineer at a 50-person e-commerce company
Route support emails to returns, billing or technical teams on its own GPU
One call answers the routing choice, a yes/no on receipts and an urgency score together. Its low calibration error means a rule like "auto-route above 0.9, else send to a person" behaves as set; Apache 2.0 means no licence review.
An ML engineer replacing an LLM judge in a RAG pipeline
Score whether each retrieved passage answers the question
A yes/no question per passage returns a probability without generating text, on one GPU. Measure accuracy on your own labelled set first: on the community index it is right about 66% of the time across all areas, and 58% on knowledge questions.
Before you commit
- Setup
- pip install transformers 5.17 or later with torch and safetensors, load with trust_remote_code=True and call model.system_one(state, questions); a transformers "decision" pipeline is also registered.
- Integrations
- Questions use the System One schema (noul, choice, score) shared by Jev, pplx-decider and Liquid's d1, so harnesses written for those work unchanged. It is the decision model family of the vLLM Semantic Router project.
- Data and privacy
- Self-hosted only, so inputs stay on your hardware. Apache 2.0 permits commercial use without a revenue cap.
- Cost at scale
- No API price; 4.21 billion parameters run on a single GPU, and the community index ran it on one RTX PRO 6000 at a 33.5 ms median. Several questions about one input share a single forward pass.
- Limits
- 16,384-token context, text and JSON only, three answer types (no numbers, boxes or extraction). Weakest on knowledge (58% accuracy) and arts (50%) areas of the community index.
What’s good
- Expected calibration error 0.030 on the community index
- Apache 2.0 with no revenue limit
- Ahead of Decider 4B and JPT-4B on the community index
- Training data audited against index test items, per the card
What’s not
- 9 points behind the team's own d3-mini (49.97)
- Text only; no images
- 33.5 ms median on the community run, not the card's 12.9 ms
- No hosted API
Best for
Engineers who need a small, self-hosted, permissively licensed text classifier whose probabilities can be trusted for thresholds.
Decision-2.0-Nox-4B (vLLM Semantic Router) alternatives
Decision-2.0-Nox-4B (vLLM Semantic Router): common questions
- Is Decision-2.0-Nox-4B (vLLM Semantic Router) worth it?
- Honest numbers and good calibration under Apache 2.0, but at 40.75 on the community index it trails the team's own d3-mini by nine points. We rate it 3.4 out of 5.
- How much does Decision-2.0-Nox-4B (vLLM Semantic Router) cost?
- Free weights under Apache 2.0; no hosted endpoint for this model.
- What are the downsides of Decision-2.0-Nox-4B (vLLM Semantic Router)?
- 9 points behind the team's own d3-mini (49.97); text only; no images; 33.5 ms median on the community run, not the card's 12.9 ms; no hosted API.
- Who is Decision-2.0-Nox-4B (vLLM Semantic Router) best for?
- Engineers who need a small, self-hosted, permissively licensed text classifier whose probabilities can be trusted for thresholds.
- How much does Decision-2.0-Nox-4B (vLLM Semantic Router) cost for a team or in production?
- No API price; 4.21 billion parameters run on a single GPU, and the community index ran it on one RTX PRO 6000 at a 33.5 ms median. Several questions about one input share a single forward pass.
- Is Decision-2.0-Nox-4B (vLLM Semantic Router) safe for company data?
- Self-hosted only, so inputs stay on your hardware. Apache 2.0 permits commercial use without a revenue cap.
- What are the best Decision-2.0-Nox-4B (vLLM Semantic Router) alternatives?
- The closest alternatives we have reviewed are Perplexity Decider (pplx-decider), d3-turbo (vLLM Semantic Router), d3 (vLLM Semantic Router), LFM3 and Jev.
Answers checked Oct 31, 2026.