Diffusion Jev (DiffusionGemma 26B-A4B on patched SGLang, 48-step diffusion readout, self-hosted)v1.7 · 10 s rule

Data: JevBench v1.8.18, published 2026-11-20 · method v1.7 (10 s rule), re-scored from the recorded per-decision times

Capability 58.1 · rank #153 of 251 on the open-weights board

Last measured: 2026-10-06 · measurement release: v1.6.1

apache-2.0 (model card metadata); linked Gemma 4 model terms

Base model: google/diffusiongemma-26B-A4B-itsource

intelligence

54.9

calibration

61.3

speed

86.3

cost

71.2

Composite (secondary)

66.5 · Composite rank #104

Cost and measurement conditions

$0.0091 per 1,000 decisions (estimated)

GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: moe gemma, 4B active parameters; throughput 34662.5 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 763,351 input + 7,370 output tokens over 1,477 decisions = 521.815165877 tokens/decision. GPU-time cost USD 0.00910120155482596/1,000; old list-price cost USD 0.0480112322274881/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00910120155482596/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases.

Measured p50 latency: 0.131 s (on-GPU, no network, loopback; not directly comparable with API latency, which we time end to end from our measuring host, network included)

Ranking and cap use: 0.41 s (est. incl. network (2x + 150 ms))

Self-hosted model: measured p50 0.131 s on-GPU, no network (timed on our GPU over loopback). Not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and latency cap use the same estimate for every self-hosted row: 2 × measured + 150 ms = 0.41 s. Model authors can ask us to time a hosted endpoint instead (on request).

x2 + 0.15 s (assumption, not measured)

measured 2026-10-06 on the live v1.6.0/v1.6.1 pool. github.com/Hangzhi/diffusion-jev-sglang@d61f0a9c0c9b6d78631aa250125dc4b20e9ea0a2 serving google/diffusiongemma-26B-A4B-it@f7f5b7f5fa82ffc52addd066915886d497f5517b on SGLang@ef90bf677e4e6b280bf796f1c5f4ef0f05dda27f (their open PR 34061) with their own version-checked readout and context-admission patch. Built from THEIR OWN tested container recipe (deployment/modal_app.py + deployment/gemma-requirements.txt), which their doc names as the recipe to use; their `engine` extra and top-level Dockerfile target the older LLaDA deployment and their doc says plainly not to use them as a DiffusionGemma installer, so they were not used. SGLang is a pure-python install imported through PYTHONPATH, no source compile. Served with their own scripts/launch_diffusiongemma.py: native Gemma4Renoise, a 256-position canvas, up to 48 adaptive denoising steps, seed 42, thinking disabled, 8,192-token context, Triton attention, no CUDA graphs, no prefix cache, T=1 softmax over the options offered - their configuration, unchanged, on one H100 80 GB (Lium) against the A100 80 GB they tested. THE AUTHOR'S OWN STATEMENT, CARRIED UNSOFTENED: these are self-conditioned denoiser scores, NOT independent masked-token likelihoods and NOT calibrated probabilities of correctness; this is not a one-forward-pass or zero-generation method. There are no new trained weights - it is an inference configuration over Google's unchanged checkpoint. A format mismatch returns their own HTTP 503 and is recorded invalid, the same way for everyone. Latency axis uses the standard self-hosted x2 + 0.15 s adjustment. add-requests run 55, GitHub #30

Measured latency over 1,500 decisions: p50 0.131 s · p95 0.210 s

0 decisions over 10 s.

A decision that takes longer than 10 seconds (measured time) counts as a wrong answer and as zero calibration for that item: each type's calibration is multiplied by the share of its decisions answered within 10 s. The limit uses the measured time (raw). The displayed adjusted time for self-hosted rows (2 × + 0.15 s) is an assumption and is not used for the limit.

Breakdowns (not part of the score)

Chance-corrected competence (0 = chance, 100 = perfect) by category on S+P+L1+L2+L3. Raw values, not equated; the same cells feed the compare view.

Capability by subject topic

Rules, policy & law52.2n=1616
Coding & software73.3n=382
Math & numbers16.7n=370
Finance & commerce54.3n=223
Support & operations55.6n=160
Everyday language74.2n=145
Safety & security73.0n=109

Use cases (TypeSafe categories)

Model routing77.8n=524
Legal & compliance57.1n=515
Customer support53.4n=240
E-commerce47.9n=119
Insurance claims50.6n=119
Risk assessment54.5n=118
Financial crime30.2n=101
Feature extraction52.3n=99
Lead generation52.9n=95
Recruiting32.3n=94
Knowledge graphs19.4n=94
LLM guardrails78.9n=93
Moderation62.2n=93
Code linting38.8n=90
Search & retrieval75.8n=90
Science29.8n=90
Advertising30.9n=88
Gaming23.2n=86
Demand forecasting13.7n=83

Cells need at least 30 answered items; † fewer than 30 (indicative only; drawn as hollow, lighter points on the compare radars).

JevBench v1.6.43 values (method v1.6, before the 10 s rule)

Capability 58.09 · Composite 59.33 · rank #58 on the all-systems board · #34 open weights

intelligence 54.86 · calibration 61.31 · speed 86.79 · cost 49.56

v1.6 speed basis (124 items): p50 0.131 s · p95 0.179 s

These figures are kept for comparison. The current figures above are re-scored under the 10 s rule from the recorded per-decision times and replace them.

Published source