What elephant-1 is
elephant-1 is the smaller of two decision models that Base Labs, Baseten’s research team, released on 21 October 2026. Its sibling is the 27B mammoth-1. Base Labs post-trained Alibaba’s Qwen3.5-9B on 6.21 million decision examples and used this model to test each data mix before training the larger one. Baseten says to pick elephant-1 for quick decisions and mammoth-1 when a task needs more general knowledge.
It answers Choice, Score and Noul questions about text and images, returns a probability for every option, and writes no text.
Access and price
Baseten serves it at POST https://inference.baseten.co/v1/systemone, the same request shape as Jev, with the model id decision-models/baseten-elephant-9b-latest. It is free for one week on Baseten’s Model APIs, until 27 October 2026, and OpenRouter lists a free version served by Baseten. Baseten has not published a price for after the free week. A request can hold 16,384 tokens; longer ones are refused. The weights are on Hugging Face under MIT.
What the benchmarks show
On the Artificial Analysis Decision Index v1.0, checked on 23 October 2026, Elephant 9B scores 64.69. That puts it 3rd of 17 decision models, behind d1-ultra at 65.25 and SubQ Snap 1.1 at 65.00, and ahead of Jev 1.14 at 57.07. Artificial Analysis measured a median of 165 ms per task through Baseten’s API. Its weakest part is knowledge: 11.6 on Omniscience accuracy and 18.3 on HLE multiple choice.
On the Jev Decision Index 0.4.1, generated on 22 October 2026, it scores 60.26 on the Full score, 40th of 229 rows and level with Jev 1.14 at 60.24. Its public score of 72.24, 4th on the board, comes from a run Base Labs submitted. Its calibration error is 0.0578 against Jev 1.14’s 0.0799.
Baseten’s blog calls it tied 8th of 208 at 62.65. That figure comes from the board’s earlier 0.4 edition.
Things to check
The Index row records that elephant-1’s training included the train splits of Decision Index benchmarks, as the authors declared. The Index also found signs of overlap with one private new-domain test and left that test out of its score. Base Labs’ report says it removed examples that overlapped either index. The gap between its public score and its Full score is 12 points, so test it on your own data.
Specifications
| System One Score | 91.7 from 2 of 3 boardsArtificial Analysis Decision Index 99.2, Jev Decision Index 84.3. How the score works |
| Question types | ChoiceScoreNoul |
| Max Choice options | Not documented |
| Score levels | Not documented |
| Questions per call | Not documented |
| Total context | 16,384 tokens |
| State budget | Not documented |
| Rate limit | Baseten publishes no rate limit for the model as of 2026-10-23. |
| Endpoint | POST https://inference.baseten.co/v1/systemone |
| Licence | MIT |
| SDKs |
The model card gives a maximum sequence length of 16,384 tokens and says longer requests are refused, not truncated. OpenRouter lists a 16,384-token context. The card and Baseten's library page say it reads text and images; OpenRouter lists text input only. Self-hosting needs one CUDA GPU and the pinned runtime that setup.sh installs.
Versions
- decision-models/baseten-elephant-9b-latest, 21 Oct 2026, The model id on Baseten's Model APIs. OpenRouter lists it as baseten/elephant-1:free, snapshot baseten/elephant-1-20261021. Weights on Hugging Face as baselabs/elephant-1, a full fine-tune of Qwen3.5-9B with 9.4B parameters. The Jev Decision Index scored revision 14537f0. Release notes
Use cases
What people use elephant-1 for, one page per pattern.
Support inbox triage with System One models
Send a support ticket to Jev once with every question attached. Category comes back as a selected label, severity and frustration as numbers on scales you wrote, refund intent as a probability. Your code reads those values and decides what happens to the ticket.
Intent and model routing with System One models
One Jev call reads an incoming request and returns its intent as a label plus a difficulty rating on a scale you wrote. Your router reads both numbers and picks the handler: deterministic code, a cheap model, an expensive one, or a human queue.
LLM guardrails with System One models
Put one Jev request in front of an LLM and one behind it. Yes/no questions return the probability that each hazard holds, a Score rates how much harm complying would do, and your thresholds turn those numbers into pass, review, block, or a crisis path.
Confidence-gated actions with System One models
Jev returns a confidence value from 0 to 1 alongside every Choice and Score answer. Your code treats it as a separate axis: act automatically when it's high, confirm or flag when it's middling, hand the decision to a person when it's low. Riskier actions get higher bars.
Benchmarks
Published scores, each linked to its source. A row marked vendor-run was measured by the vendor itself, so read it as one side's result. The compare hub sets these beside the other models.
| Benchmark | Metric | Score | Run by | Date |
|---|---|---|---|---|
| Artificial Analysis Decision Index 1.0 | Index score | 64.69 score | Independent | 23 Oct 2026 |
| Artificial Analysis Decision Index 1.0 | Expected calibration error | 0.038 ECE | Independent | 23 Oct 2026 |
| Artificial Analysis Decision Index 1.0 | Cost per 1M tasks | 8.4 USD | Independent | 23 Oct 2026 |
| Artificial Analysis Decision Index 1.0 | Time per task | 165 ms | Independent | 23 Oct 2026 |
| Jev Decision Index 0.4.1 | Full score | 60.26 score | Independent | 22 Oct 2026 |
| Jev Decision Index 0.4.1 | Public score | 72.24 score | vendor-run | 22 Oct 2026 |
| Jev Decision Index 0.4.1 | Expected calibration error | 0.0578 ECE | Independent | 22 Oct 2026 |
- Artificial Analysis Decision Index, Index score: Row: Elephant 9B, hosted by Baseten, 3rd of 17 decision models behind d1-ultra (65.25) and SubQ Snap 1.1 (65.00). Sub-scores: Omniscience accuracy 11.6, non-hallucination 93.2, HLE-MultiChoice 18.3, When2Call 94.8, WildGuardTest F1 93.3. Jev 1.14 scores 57.07.
- Artificial Analysis Decision Index, Expected calibration error: Row: Elephant 9B. Computed by this site from AA's published calibration bins (all answers pooled); AA does not publish an ECE. Lower is better.
- Artificial Analysis Decision Index, Cost per 1M tasks: Row: Elephant 9B. AA's measured cost on its four evals. Baseten has not published a list price.
- Artificial Analysis Decision Index, Time per task: Row: Elephant 9B. Median time per task.
- Jev Decision Index, Full score: Row: baselabs elephant-1 (9B), 40th of 229 rows by this score. Counts 60% private same-skill tests and 40% private new-domain tests; public benchmarks no longer count. 0 is random guessing, 100 is perfect. The Index ran the private tests on one RTX PRO 6000 with Base Labs' serving code. The row notes that training included the train splits of Decision Index benchmarks (author-declared), and the Index left out one private new-domain test after finding signs of overlap. Jev 1.14 scores 60.24.
- Jev Decision Index, Public score: Row: baselabs elephant-1 (9B), 4th of 229 rows by this score. Base Labs submitted the public run, and the Index rescored it with its own kit. Jev 1.14 scores 60.7.
- Jev Decision Index, Expected calibration error: Row: baselabs elephant-1 (9B). Lower is better. Jev 1.14 scores 0.0799.