MITdecisionRuns locally27B (mammoth-1); 9B (elephant-1 sibling)
Mammoth-1: Specs, License & How to Run
Base Labs MIT 27B decision model (9B Elephant-1 sibling) for System-One routing and classification.
- License
- MIT
- Parameters
- 27B (mammoth-1); 9B (elephant-1 sibling)
- HF likes
- 44 (2026-10-23)
- First seen
- 2026-10-19
TL;DR. Mammoth-1 is a decision model from Base Labs (Baseten). Parameters: 27B (mammoth-1); 9B (elephant-1 sibling).License: MIT. It is listed under LLM classification APIs.
| Developer | Base Labs (Baseten) |
|---|---|
| Type | decision |
| Architecture | Decision models post-trained from Qwen3.5-27B (mammoth-1) and Qwen3.5-9B (elephant-1). HF model_type qwen3_5; pipeline_tag text-classification; tags include decision-model, classification, calibration, routing. Vision encoder frozen; block-bidirectional mask on full-attention layers (card). |
| Parameters | 27B (mammoth-1); 9B (elephant-1 sibling) |
| Modalities | text, image |
| Languages | en |
| Context | 16,384 tokens (card) |
| License | MIT — Hugging Face cardData.license is mit and Hub tags include license:mit on both baselabs/mammoth-1 and baselabs/elephant-1 as of 2026-10-23. A raw fetch of resolve/main/LICENSE 404’d on both cards. Not legal advice. |
| Hardware | Not yet verified — Card setup.sh installs a pinned OneJev runtime (serve.patch) and asks for one CUDA GPU plus pytorch/pytorch:2.8.0-cuda12.8-cudnn9-devel. HF usedStorage is 109,447,565,536 bytes on mammoth-1 (~101.9 GiB) and 18,841,783,403 bytes on elephant-1 (~17.5 GiB). safetensors.parameters: 27,356,728,560 and 9,409,813,744. Peak VRAM is not independently measured. (not independently tested) |
| Hugging Face likes | 44 (as of 2026-10-23) |
| Model card | https://huggingface.co/baselabs/mammoth-1 |
| First seen | 2026-10-19 |
Mammoth-1 is a 27B decision model from Base Labs (Baseten). This page is the family entry for the paired open-weight cards baselabs/mammoth-1 and the 9B sibling baselabs/elephant-1. Official hosts we opened on 2026-10-23: both ungated cards, the technical report Mammoth-1 and Elephant-1: frontier decision models, Baseten’s Decision models docs, and the 21 October 2026 changelog elephant-1 and mammoth-1 available on Baseten. resolve/main/LICENSE 404’d on both cards.
This identification page hangs off LLM classification APIs. That mapping is editorial. We have no sourced quality comparison against Clef, Jev, or LiquidAI d1. Head-to-head facts versus the decision model already on this site: Mammoth-1 vs Cloudflare Clef.
The cards say: provide a state (text, optionally images) and a set of questions; the model returns a calibrated probability for every option from one shared-prefix prefill. That is the System One / decision-model job, not chat. Superlatives (“frontier”, “#1 of 208”) stay vendor.
What we verified (2026-10-23)
| Field | Mammoth-1 | Elephant-1 | Source |
|---|---|---|---|
| Host | huggingface.co/baselabs/mammoth-1 (not gated) | huggingface.co/baselabs/elephant-1 (not gated) | HF API |
| Created | 2026-10-19T19:35:12Z | 2026-10-19T19:35:08Z | same |
| Last modified | 2026-10-22T18:40:29Z | 2026-10-22T18:40:27Z | same |
| Likes / downloads | 44 / 99 | 34 / 85 | same |
| License tag | mit | mit | HF cardData |
| LICENSE file fetch | 404 | 404 | this check |
| Pipeline tag | text-classification | text-classification | HF API |
| Base named | Qwen3.5-27B | Qwen3.5-9B | cards |
| Params (card) | 27B | 9B | cards |
| safetensors.parameters | 27,356,728,560 | 9,409,813,744 | HF API |
| usedStorage | 109,447,565,536 (~101.9 GiB) | 18,841,783,403 (~17.5 GiB) | same |
| Max sequence | 16,384 tokens | 16,384 tokens | cards |
| Inputs | text + images | text + images | cards |
Elephant-1 (sibling, same page)
Use Mammoth-1 when you want the larger 27B card. Use Elephant-1 when you want the 9B card the family describes as faster / cheaper. Both share the same question schema (noul, choice, score), the same 16,384-token length, and the same MIT card tag. We are not creating /models/elephant-1.
How to run (from the card, not tested)
Self-hosted path copied from the mammoth-1 card:
hf download baselabs/mammoth-1 --local-dir mammoth-1 && bash mammoth-1/setup.shThe card says setup.sh installs a pinned OneJev runtime (serve.patch) and needs one CUDA GPU plus the pytorch/pytorch:2.8.0-cuda12.8-cudnn9-devel image. A Python example constructs a DecisionEngine("baselabs/mammoth-1") and a SystemOneRequest. Requests longer than the trained limits are refused, not truncated (card). forkalt has not run this.
Hosted path named in Baseten docs (21 October 2026 changelog): POST https://inference.baseten.co/v1/systemone with model slugs decision-models/baseten-mammoth-27b-latest or decision-models/baseten-elephant-9b-latest. That is a paid/hosted API after any promo window, not the weight license.
License
Card tag MIT on both siblings. A raw LICENSE fetch 404’d on 2026-10-23. Not legal advice.
What we did not invent
- Jev Decision Index 0.4 (71.24 mammoth-1 / 62.65 elephant-1) and the Artificial Analysis replica (65.9 / 65.2) stay self-reported. The cards’ comparisons to Perplexity Decider, Jev 1.14, Cloudflare Clef, and d1-ultra are upstream.
- We do not rank these weights against Clef or any closed classification API.
Related decision / classification rows: Cloudflare Clef — Mammoth-1 vs Cloudflare Clef. Chat-style open weights stay on closed LLM APIs. LiquidAI d1 remains on the classification watch list (no page).
Benchmarks
Quoted from the cited source. forkalt has not independently reproduced these numbers.
| Name | Score | Source |
|---|---|---|
| Jev Decision Index 0.4 (card table) | 71.24 (self-reported) | https://huggingface.co/baselabs/mammoth-1 |
| Artificial Analysis Decision Index replica (card table) | 65.9 (self-reported) | https://huggingface.co/baselabs/mammoth-1 |
Deploy options
Comparisons
FAQ
Is Mammoth-1 a chatbot?
No. The card describes a decision model: give it a state (text, optionally images) plus typed questions and it returns a calibrated probability for every option in one shared-prefix prefill. Question types on the card are noul (yes/no), choice, and score. It is not a free-text generator.
What is Elephant-1, and why isn’t it a second page?
elephant-1 is the 9B sibling (baselabs/elephant-1; created 2026-10-19T19:35:08Z; 34 likes / 85 downloads in the 2026-10-23 API check; MIT card tag). Same System One job, smaller checkpoint. This page is the family entry; we do not ship a thin duplicate URL.
Are the weights open, and what is the license?
Both cards are ungated. mammoth-1 (created 2026-10-19T19:35:12Z; 44 likes / 99 downloads) and elephant-1 tag license mit. resolve/main/LICENSE 404’d on both in this check. Not legal advice.
How does this differ from Cloudflare Clef?
Same product class (typed options, not chat). Mammoth-1/Elephant-1 are MIT-tagged Qwen3.5 post-trains with a 16,384-token card length and a System One / OneJev serve path. Clef is Apache-2.0, post-trained from Qwen3.8-27B, and also names video. See Mammoth-1 vs Cloudflare Clef. Card Decision Index rows stay self-reported.
Can I use the hosted Baseten API?
Baseten documents System One slugs decision-models/baseten-mammoth-27b-latest and decision-models/baseten-elephant-9b-latest. The mammoth-1 card (opened 2026-10-23) said a free week on Baseten Model APIs ended 2026-10-27 23:59 PT. That is a hosted SKU, not the MIT weight grant. forkalt has not called the API.
Sources & verification
- https://huggingface.co/baselabs/mammoth-1
- https://huggingface.co/baselabs/elephant-1
- https://labs.baseten.co/articles/mammoth-1-and-elephant-1-frontier-decision-models
- https://docs.baseten.co/inference/model-apis/decision-models
- https://www.baseten.co/resources/changelog/elephant-1-and-mammoth-1-available-on-baseten/
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.