djev-spark
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity2.8/10
Working Docker/vLLM path plus a Cloud Run deploy pack (djev-run); still GPU-class ops, not a laptop peer.
- Capability6/10
Decision Index 0.4.1 Full 38.15 (#126); ImageJevBench v0.3.0 calibration 47.5, p50 0.54 s. Choice / score / boolean over a DiffusionGemma 26B-A4B server; real on Spark or Cloud Run boxes, not a portable drop-in.
- Adoption32/100
Two official @googlegemma amplifications (18 Sep + 22 Sep Cloud Run) plus Spark/GB10 chatter; still niche without GPU budget.
Vendor claims
- Implements Jev POST /v1/systemone in front of vLLM structured-reads on DiffusionGemma 26B-A4B NVFP4 (author)[Vendor claim — not independently verified]
- Target hardware: DGX Spark / GB10 (aarch64 unified memory, CUDA 13) (author)[Vendor claim — not independently verified]
- Cloud Run one-command deploy via taeold/djev-run; tweet cites ~35–60 ms single-step and ~$3/hr (UnverifiedClaim)[Vendor claim — not independently verified]
News covering this model
publicado (OK do Fabricio 17/10 BRT). Board audit of 17 Oct 2026: see the “Board audit” section below; where it conflicts with older board statements on this page, it wins.
djev-spark is a hardware-oriented System One stack: DiffusionGemma 26B-A4B NVFP4 served through a structured decision server that speaks Jev’s POST /v1/systemone — by mmastrac, not TypeSafe.
Primary path: DGX Spark / GB10 class boxes (github.com/mmastrac/djev-spark). Deploy pack: taeold/djev-run wraps that stack for Google Cloud Run (RTX PRO 6000) with built-in /snake and /dino demos.
Awareness: @googlegemma amplified “DiffusionGemma as Jev” (18 Sep); follow-up Cloud Run post (22 Sep). Treat as adoption signal — speed/pricing/JevBench numbers remain UnverifiedClaim. See news and Cloud Run follow-up.
Related deploy / demos
| Project | Role |
|---|---|
| taeold/djev-run | Cloud Run GPU image + one-command deploy of a Jev-compatible /v1/systemone endpoint; browser Snake/Dino UIs. Latency, $/hr tables in-repo = UnverifiedClaim |
JevBench v1.4.2
| Metric | Score | Rank |
|---|---|---|
| Overall | 52.2 | #6 |
| Intelligence | 45.3 | - |
| Calibration | 72.8 | - |
| Speed | 78.4 | - |
| Cost | 54.1 | - |
The DiffusionGemma backbone punches above its weight on decision quality — #6 overall puts djev-spark ahead of many larger models. The speed score (78.4) reflects Cloud Run latency; dedicated Spark hardware would score higher.
Limits (read before you ship)
- Still not a “pip install on a laptop” peer to Kev / Laya; Cloud Run removes owning a Spark box but not GPU-class cost/ops.
- Forked vLLM + overlay patches — expect ops drift as nightlies move.
- Catalog scores reflect “interesting specialized stack,” not general availability.
Why it is in the catalog
Tracks a real System One API on a diffusion MoE backbone in the S1Bench gap set — useful signal for Spark / Cloud Run builders; keep it distinct from portable open weights like Decider.
Board audit (17 Oct 2026)
publicado (OK do Fabricio 17/10 BRT). Checked against the official boards on 17 Oct 2026: Decision Index 0.3.2 (data of 16 Oct, 19:00 UTC), JevBench and ImageJevBench live. The boards are separate sources and are never averaged. Older board statements and rankings further down this page are superseded where they conflict with this section.
| Board | Row | Result |
|---|---|---|
| Decision Index 0.3.2 (16 Oct 19:00 UTC) | djev diffgemma | Full 41.78 (#89, tie group 21); public 40.89 (#97); same-skills 41.63; new-domains 38.59 |
| ImageJevBench v0.3.0 | djev-spark NVFP4 | Capability 24.5; Intelligence 1.5; calibration 47.5; p50 0.54 s; $0.064/1k |
| ImageJevBench v0.1.5 archive (frozen) | djev-spark NVFP4 | Capability 66.7; Intelligence 49.1; calibration 84.3; p50 0.50 s; $0.063/1k |
| ImageJevBench v0.3.0 | djev-dev BF16 | Capability 25.4; Intelligence 3.9; calibration 46.9; p50 0.59 s; $0.072/1k |
| ImageJevBench v0.1.5 archive (frozen) | djev-dev BF16 | Capability 63.8; Intelligence 49.6; calibration 78.0; p50 0.54 s; $0.070/1k |
| ImageJevBench v0.3.0 | djev-distill-v4 | Capability 40.2; Intelligence 18.8; calibration 61.7; p50 0.61 s; $0.072/1k |
| ImageJevBench v0.1.5 archive (frozen) | djev-distill-v4 | Capability 68.7; Intelligence 52.5; calibration 84.8; p50 0.54 s; $0.068/1k |
Score proposal (from scratch, audit follow-up 17 Oct 2026; publicado (OK do Fabricio 17/10 BRT)). Capability 6.8 → 6.2 = types × 0.30 + calibration × 0.25 + latency × 0.25 + accuracy × 0.20 = 10 × 0.30 + 3 × 0.25 + 5 × 0.25 + 6 × 0.20 = 6.20 → 6.2. Types 10 (choice, score, boolean). Calibration 3: no author ECE and no text-board axis; the ImageJevBench v0.3.0 calibration for djev-spark NVFP4 is 47.5 (below 60 → 3). Latency 5: the same board’s p50 is 0.54 s (>0.5 s and ≤1 s → 5); the ~35–60 ms tweet is UnverifiedClaim. Accuracy 6: Decision Index Full 41.78 ÷ 60.11 × 8 = 5.6 → 6. The old ‘JevBench #6 (52.2)’ is from v1.4.2; the frozen ImageJevBench v0.1.5 archive row (capability 66.7) is not used. EN and pt-br were out of sync (EN 6.8, pt-br 5.4); both now carry the proposal. Maturity and adoption unchanged.
Board audit (20 Oct 2026, DI 0.4)
publicado (OK do Fabricio 20/10 BRT).
Decision Index 0.4 (20 Oct; Full now from the private parts only) lists djev diffgemma at Full 39.18 (#108); it was 41.78 on the archived 0.3.2 edition. Accuracy rule, reference Jev 1.13 on DI 0.4 Full (61.26): 8 × 39.18 ÷ 61.26 = 5.12 → 5 (was 6). Capability = 10 × 0.30 + 3 × 0.25 + 5 × 0.25 + 5 × 0.20 = 6.00 → 6.0 (was 6.2). Other sub-scores, maturity and adoption unchanged.
Board audit (22 Oct 2026, DI 0.4.1)
publicado (OK do Fabricio 22/10 BRT).
Decision Index 0.4.1 (22 Oct; reference Jev 1.13 Full 59.22, was 61.26 on 0.4) lists this page’s row at Full 38.15 (#126). Accuracy rule: 8 × 38.15 ÷ 59.22 = 5.15 → 5 (was 5). Capability unchanged. Maturity and adoption unchanged.