Benchmarks
More
OverviewBenchmaxxingCompareChartsCost vs CapabilityProviders per ModelProvider explorerGatewaysEU & SovereignJev-class modelsImage JevBenchAbout
JevBench by Benchmark Heaven · v1.4.2.2 · individual system

Needle 3 (Cactus, 2-bit, local CPU)

Small tool-calling model · by Cactus Compute

JevBench v1.4.2.2 score

—

Partial run, not ranked — No completed sealed v1.4 run; retained as an unranked historical row.

Shown, not ranked. The marked tick is Jev 1.13.0 (63.3).

Published axes

intelligence
—
calibration
—
speed
—
cost
—

Against Jev 1.13.0

Four radars compare this fixed pair across the score axes, accuracy per tier, and accuracy by family on the hard tier and sealed set. Further out is better on every spoke.

  • A: Needle 3 — Small tool-calling model · Score — (partial run, not ranked)
  • B: Jev 1.13.0 — Jev (TypeSafe, closed) · Score 63.3 (#4)

The four score axes

Radar: the four score axes, two systemsThe four score axes, Needle 3 vs Jev 1.13.0. Intelligence: none (0) vs 53.1; Calibration: none (0) vs 76.3; Speed: none (0) vs 83.3; Cost: none (0) vs 52.0.50100Intelligencenone (0) · 53.1Calibrationnone (0) · 76.3Speednone (0) · 83.3Costnone (0) · 52.0
0–100, the values in the table. A label-only system has no calibration (counted as 0).

Accuracy per tier, incl. sealed

Radar: accuracy per tier, incl. sealed, two systemsAccuracy per tier, incl. sealed, Needle 3 vs Jev 1.13.0. Easy: 47% vs 100%; Standard: 17% vs 99%; Judge: 32% vs 95%; Hard: 8% vs 74%; Sealed: — vs 37%.50100Easy47% · 100%Standard17% · 99%Judge32% · 95%Hard8% · 74%Sealed37%
Share correct per tier; Sealed = the 308 private decisions, aggregate only.

Current question set by family (hard + sealed)

Needle 3 was not run on the full v1.4 question set (a partial run); families without both hard-tier and sealed results are left out (—).

Radar: current question set by family (hard + sealed), two systemsCurrent question set by family (hard + sealed), Needle 3 vs Jev 1.13.0. Ambiguous / abstain: — vs 43%; Judge: — vs 54%; Long policy: — vs 44%; Multi-hop: — vs 64%; Probability: — vs 63%; Temporal / numeric: — vs 28%; Trade-off: — vs 55%; Routing: 0% vs 100%; Trap / adversarial: — vs 83%; Paraphrase: — vs 64%; Safety judge: — vs 38%.50100Ambiguous /abstain43%Judge54%Long policy44%Multi-hop64%Probability63%Temporal /numeric28%Trade-off55%Routing0% · 100%Trap /adversarial83%Paraphrase64%Safety judge38%
Share correct per family across the 220 hard-tier decisions (public and held out) and the 308 sealed decisions of v1.4, pooled; Routing is hard-tier only, Paraphrase and Safety judge sealed only.

Sealed set by family

Needle 3 has no sealed family breakdown.

Radar: sealed set by family, two systemsSealed set by family, Needle 3 vs Jev 1.13.0. Ambiguous / abstain: — vs 30%; Judge: — vs 34%; Long policy: — vs 28%; Multi-hop: — vs 45%; Paraphrase: — vs 64%; Probability: — vs 50%; Safety judge: — vs 38%; Temporal / numeric: — vs 29%; Trade-off: — vs 38%; Trap / adversarial: — vs 42%.50100Ambiguous /abstain30%Judge34%Long policy28%Multi-hop45%Paraphrase64%Probability50%Safety judge38%Temporal /numeric29%Trade-off38%Trap /adversarial42%
Share correct within each sealed family — system-level aggregates; the items stay private.

Availability and evidence

Openness
Code and weights marked open in the published row
License note
Apache-2.0 (model and package)
Cost evidence
estimated; the board’s row disclosure contains the published basis.
Endpoint condition
our CPU (2 threads, Ryzen 5 3600)
Note on this row
V1.4: Not re-measured: ~100-250 s per item on a rented CPU pod (19 s on Sandy); needs a dedicated CPU host.
Published source
https://github.com/cactus-compute/needle

From the public v1.4.2.2 aggregate. Scores and ranks can change when a new release is published.

Read the full board and published method. The overall score is a composite, not raw accuracy.