LangWatch
Sign Up

vLLM Semantic RouterOpen weights model4B parameters

Decision 2.0 Nox 4B benchmarks

All models
  • General rank
    Too few tasks3 non-coding tasks ranked5 needed for a rank
  • Speed#7 / 8
    691 msMedian latency, p50L4 only
  • Cost
    Self-hostedRan on L4No hosted price
  • LLM-as-a-judge#2 / 31 tied
    61.3%Best task · Accuracyvs Jev-class models

Comparison summary

Strongest
Judging and evals
Leads 0 of 3 (0%)

A typed decision model: it answers a typed question, such as yes or no or one of a few options, with a probability for each answer and no text. Strongest at judging.

Specifications

Lab
vLLM Semantic Router
Weights
Open weights
Parameters
4B
Licence
Apache-2.0

Judging and evals

Set

Share of tasks led · #23 of 31

  1. Mercury DecideLeads 2 of 3
  2. OpenAI Decisions APILeads 4 of 6
  3. ClefLeads 1 of 3
  4. deck-31BLeads 1 of 3
  5. Kev-27BLeads 1 of 3
  6. pplx-decider v1.1 27BLeads 1 of 3
  7. Quyet-1.0-LargeLeads 1 of 3
  8. Decision 2.0 Nox 4B
    Leads 0 of 3
See the 3 tasks

LLM-as-a-judge

Tier 2 of 2#2 / 31

Accuracy, % · Higher is better

  1. OpenAI Decisions API71.7, interval 66.1 to 77.2, ahead alone
  2. Clef Flash63.0, interval 56.2 to 69.8, tier 2
  3. Mercury Decide61.7, interval 55.6 to 67.9, tier 2
  4. Decision 2.0 Eos 0.8B61.3, interval 54.6 to 68.0, tier 2
  5. Decision 2.0 Nox 4B
    61.3, interval 55.2 to 67.7, tier 2
  6. Decision 2.0 Sol 2B61.3, interval 54.4 to 67.9, tier 2
  7. Jev61.3, interval 54.8 to 67.9, tier 2
  8. pplx-decider v1.1 27B61.3, interval 55.8 to 67.1, tier 2
Tested against Jev and 29 Jev-class models

Scenario judge

Tier 3 of 4#19 / 31

Accuracy, % · Higher is better

  1. Mercury Decide83.3, interval 75.5 to 90.3, tied for the lead
  2. OpenAI Decisions API83.3, interval 76.1 to 89.9, tied for the lead
  3. deck-31B81.9, interval 74.8 to 88.1, tied for the lead
  4. Quyet-1.0-Large79.2, interval 72.0 to 85.9, tied for the lead
  5. pplx-decider v1.1 27B77.8, interval 69.8 to 84.9, tied for the lead
  6. Clef75.7, interval 67.4 to 83.1, tied for the lead
  7. Rene-1 31B FP875.0, interval 66.9 to 82.9, tied for the lead
  8. Decision 2.0 Nox 4B
    58.3, interval 47.9 to 68.3, tier 3
Tested against Jev and 29 Jev-class models

Search

Tier 4 of 6#18 / 31

Accuracy, % · Higher is better

  1. OpenAI Decisions API70.4, interval 66.0 to 74.8, tied for the lead
  2. Mercury Decide66.3, interval 61.5 to 71.0, tied for the lead
  3. GLiDE63.9, interval 59.5 to 68.3, tier 2
  4. Jev63.3, interval 58.2 to 68.1, tier 2
  5. Kev-27B61.1, interval 56.4 to 65.7, tier 2
  6. Clef60.2, interval 55.7 to 64.9, tier 2
  7. deck-31B58.9, interval 54.3 to 63.5, tier 2
  8. Decision 2.0 Nox 4B
    44.8, interval 39.3 to 50.2, tier 4
Tested against Jev and 29 Jev-class models

Speed

Median latency, p50 · L4 only · lower is better

  1. Decision 2.0 Kai 0.6B114 ms
  2. Decision 2.0 Eos 0.8B250 ms
  3. Decision 2.0 Sol 2B269 ms
  4. decider-4b v2409 ms
  5. Lev 4B620 ms
  6. JevK5665 ms
  7. Decision 2.0 Nox 4B
    691 ms
  8. Winnow1,356 ms
How we count

A model leads a task when it is in the task's leading tie tier. Each task counts inside its own benchmark, against that benchmark's models; no score is averaged. An area chart shows the models ranked on at least 2 of the area's tasks in a benchmark this model is in too. The overall standing counts every non-coding tasks and needs 5 for a rank. Latency is compared only on one hardware tier.

Decision 2.0 Nox 4BOther modelsWhisker: 95% interval