mammoth-1 (27B) + elephant-1 (9B)
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity7/10
MIT weights on Hugging Face (opened 21 Oct), a hosted /v1/systemone endpoint on Baseten Model APIs and Vercel AI Gateway (AI SDK, Decisions API, TypeSafe-compatible route), card plus technical report. Free until 27 Oct; later price not published. No model SLA.
- Capability8.1/10
Noul / Choice / Score from one shared prefill. Decision Index 0.4 Full 71.24 (#1, earlier weights); 0.4.1 rescored the public weights at 68.44 (tied #3; Jev 1.13 59.22). DI ECE 0.041, median 220 ms on one RTX PRO 6000. Trained with a Brier term and an "I don't know" option.
- Adoption12/100
Baseten launch post ~300 likes / 29k views, Vercel AI Gateway changelog, about 180 HF downloads and 64 likes across both repos in a day. Free week on Baseten and Vercel. No named production users yet.
Vendor claims
- mammoth-1 is the best decision model in the world, #1 of 208 on Jev Decision Index 0.4 (71.24)[Vendor claim — not independently verified]
- elephant-1 ties d1-ultra for #1 on the Artificial Analysis Decision Index (65); replica scores 65.9 (mammoth-1) and 65.2 (elephant-1)[Vendor claim — not independently verified]
- 6.21M training examples, zero contamination against the Decision Index and AA benchmarks[Vendor claim — not independently verified]
News covering this model
publicado (OK do Fabricio 22/10 BRT). Scores are calculated from scratch.
Read this first.
- One page, two models.
mammoth-1(27B) is the headline and the scored model;elephant-1(9B) has its own numbers below.- Two weight revisions. Decision Index 0.4 (20 Oct) scored an earlier revision of mammoth-1 at 71.24 (#1). After the public release the board’s 0.4.1 edition rescored the current
mainweights at 68.44 (tied #3). This page uses the row of the weights you can download.- The Artificial Analysis “tie for #1” is close but not a tie. AA’s page lists Elephant 9B at 64.7, second to d1-ultra at 65.2 (read 22 Oct). mammoth-1 is not on AA.
What it is
mammoth-1 and elephant-1 are full fine-tunes of Qwen3.5-27B and Qwen3.5-9B by Base Labs, the research arm of Baseten. Given a state and a set of typed questions (yes/no, choice, score), they return a probability for every option and generate no text. All questions share one prefill: the full-attention layers use a block-bidirectional mask, so the prefix cache is reused and each question sees its own options. The vision encoder is frozen; training was text only.
On 20 Oct the repositories were gated “for Decision Index evaluation only”, which kept them out of the catalog (news). On 21 Oct Base Labs switched both to MIT and Baseten launched them on its Model APIs.
Specs
| Attribute | mammoth-1 |
elephant-1 |
|---|---|---|
| Developer | Base Labs (Baseten) | same |
| Base | Qwen3.5-27B, full fine-tune | Qwen3.5-9B, full fine-tune |
| License | MIT | MIT |
| Types | noul, choice, score; trained “I don’t know” option | same |
| Context | 16,384 tokens; longer requests are refused, not truncated | same |
| Training | Mix 8, 6.21M examples; cross-entropy + Brier (λ = 1); 1 epoch; 72 GB300 | same data; 128 H100 |
| Self-host | setup.sh (OneJev + serve.patch), one CUDA GPU |
same |
| Hosted | Baseten POST /v1/systemone, model decision-models/baseten-mammoth-27b-latest; Vercel AI Gateway baseten/mammoth-1 |
baseten/elephant-1 |
| Price | free until 27 Oct 2026, 11:59 PM PT; later price not published | same |
Board results
| Board | Row | Result |
|---|---|---|
| Decision Index 0.4.1 (generated 22 Oct 05:40 UTC) | baselabs mammoth-1 (27B), current weights | Full 68.44, tied #3 of 225; public 75.34; ECE 0.041; median 220 ms |
| same | baselabs elephant-1 (9B) | Full 60.26 (#40); public 72.24; ECE 0.058; median 114 ms |
| same | Jev 1.13 (reference) | Full 59.22 |
| Decision Index 0.4 (20 Oct, archived edition) | mammoth-1, earlier revision | Full 71.24, #1 |
| Artificial Analysis Decision Index (read 22 Oct) | Elephant 9B | 64.7, #2 of 16 (d1-ultra 65.2; Jev 1.13 57.5) |
| JevBench (v1.6.38) | — | not listed |
Boards are read separately and never combined.
Vendor claims (UnverifiedClaim)
- mammoth-1 “#1 of 208 on Jev Decision Index 0.4” with 71.24 (report). True for the 0.4 edition and the revision scored there; the public weights are 68.44 on 0.4.1.
- elephant-1 “tied for #1” on Artificial Analysis (65 vs d1-ultra 65) (card, Vercel). AA’s page shows 64.7 against 65.2.
- Replica of the AA index: mammoth-1 65.9, elephant-1 65.2; zero contamination with both indexes (report).
Scores (from scratch, 22 Oct 2026)
Scored on mammoth-1.
- Capability 8.1 = types × 0.30 + calibration × 0.25 + latency × 0.25 + accuracy × 0.20 = 10 × 0.30 + 7 × 0.25 + 7 × 0.25 + 8 × 0.20 = 8.10. Accuracy: Decision Index 0.4.1 Full 68.44 against Jev 1.13 at 59.22 in the same edition → 8 × 68.44 ÷ 59.22 = 9.25, capped at 8 (tied #3, one board). Calibration 7: method published (Brier term, abstention training) and board ECE 0.041. Latency 7: board median 220 ms.
- Maturity 7.0 = availability × 0.30 + docs × 0.25 + integrations × 0.25 + support × 0.20 = 8 × 0.30 + 7 × 0.25 + 8 × 0.25 + 4 × 0.20 = 6.95 → 7.0.
- Adoption 12 = engagement × 0.35 + ecosystem × 0.35 + production × 0.30 = 20 × 0.35 + 12 × 0.35 + 2 × 0.30 = 11.8 → 12.
Community GGUF: elephant-1-decision-GGUF (25 Oct 2026)
publicado (OK do Fabricio 25/10 BRT).
waseigo/elephant-1-decision-GGUF (MIT, uploaded 24 Oct 23:03 BRT, corrected 25 Oct) takes the Q4_K_M elephant-1 files from mradermacher/elephant-1-GGUF and adds only the llama.cpp decision metadata (qwen35.decision.type = openjev, non-causal attention) and a systemone prompt template, so llama.cpp’s POST /v1/systemone works. No weights changed. The first upload had a broken template (the model saw no option letters); the author replaced it on 25 Oct. Not from Base Labs; scores unchanged.
joey-1 (2B), 29 Oct 2026
publicado (OK do Fabricio 30/10 BRT).
Base Labs added joey-1 on 29 Oct: Qwen3.5-2B post-trained on the same data and recipe, MIT weights, text and images, and Baseten Model APIs at $0.01 per million input tokens. Baseten says that is a tenth of mammoth-1’s price, which implies about $0.10/M for the 27B after the free week (not confirmed on a price page).
| Board | Row | Result |
|---|---|---|
| Decision Index 0.4.1 (data 29 Oct 14:04 UTC) | baselabs joey-1 (2B) | Full 35.91 (#148 of 247), public 63.30; ECE 0.044; median 85 ms |
| same | baselabs elephant-1 (9B) | Full 60.26 (#48), unchanged |
| same | baselabs mammoth-1 (27B) | Full 68.44 (#9), unchanged |
Vendor claims (UnverifiedClaim): P50 of 20 ms on a 2,000-token request, 2.5× faster than elephant-1 (49 ms); “best model under 4B on the Decision Index 0.4 public split”; 62.9 on an Artificial Analysis replica. The public-split claim matches the board, but the Full score puts joey-1 far down the table. Not on JevBench or Artificial Analysis yet.
Scores unchanged (7.0 / 8.1 / 12): this page is scored on mammoth-1. See news.