# elephant-1 (slim)

> elephant-1 is Base Labs' open-weight 9B decision model, a full fine-tune of Qwen3.5-9B that returns a probability for every option of typed questions about a state in one prefill. Base Labs is Baseten's research lab.

- Full: https://www.anchorterminal.com/tools/elephant-1.md (~9,100 tokens) · this version ~2,230 tokens · JSON https://www.anchorterminal.com/tools/elephant-1.json · canonical https://www.anchorterminal.com/tools/elephant-1
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-22 · pricing https://www.anchorterminal.com/tools/elephant-1#pricing · reviews https://www.anchorterminal.com/tools/elephant-1#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**D · 47.6/100 · rank #98 of 102 open-weight models · #21 in Decision models · not agent-ready · confidence low**


Assessment: A small cousin for routing, triage and judging, answering several typed questions about one state with probabilities in one prefill. It fits a 24 GB GPU in BF16 but runs only through a patched third-party runtime. The risks are a licence switched from proprietary to MIT on release day with no licence file, and a thin card.

## Facts

- Kind: Open-weight model · vendor: Baseten · category: Decision models · legal entity: Baseten Labs, Inc. · provenance 40/100
- Packages: huggingface `baselabs/elephant-1`
- Auth: None · pricing: Free · x402: no · licence: MIT, named only in the card's metadata (license: other, license_name: mit) since 21 October 2026, with no LICENSE file. Until then the card said proprietary, all rights reserved, access for Decision Index evaluation only. The base Qwen/Qwen3.5-9B is Apache-2.0
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/baselabs/elephant-1, read from the card and its earlier version, the files and their metadata, config.json, engine.py, setup.sh, serve.patch, the commits and discussions, Baseten's library page, blog and docs, the Base Labs report, OneJev's README and licence, the Decision Index Space's data and Artificial Analysis's index on 21 October 2026. Nothing was downloaded or run. Baseten's hosted Model API isn't graded here
- Built on: Qwen/Qwen3.5-9B (Apache-2.0), named in the card's metadata. A full fine-tune for one epoch over 6.21M decision examples with the vision encoder frozen, per the report. The chat template is byte for byte the base's
- Relation to TypeSafe's Jev: None by ownership. It answers TypeSafe's System One question types (choice, noul, score) and Baseten serves it at /v1/systemone. The Jev Decision Index is a community board in a personal Hugging Face Space that says it isn't affiliated with TypeSafe AI, and it lists Jev 1.14 among its rows
- Model: 9.41B parameters in BF16 (Hub safetensors count), 32 layers in a 3 to 1 mix of Gated DeltaNet and full attention, with the full-attention layers changed to a block-bidirectional mask so the state's cache is shared by every question. The answer is a softmax over the option letters' logits at each question's answer slot, up to 256 slots
- Files: One model.safetensors (18.82 GB, BF16), tokenizer and configs, chat_template.jinja, readout.json, and the serving bundle engine.py, setup.sh, serve.patch and constraints.txt. No pickle files. Hugging Face lists a SHA-256 for the LFS files
- Licence: MIT by the card's metadata only, since 21 October 2026, with no LICENSE file and no copyright line. The earlier card said proprietary and evaluation only. The Apache-2.0 base asks for its licence to travel with derivatives, and the repository carries no copy
- Runtimes: Only OmniJev's OneJev qev runtime at a pinned commit with the repository's serve.patch, through setup.sh. No released vLLM, SGLang, llama.cpp, Ollama or LM Studio route applies the block-bidirectional mask and slot readout, and Transformers loads the architecture but not those changes
- Context: config.json keeps the base's 262,144 positions, but the fine-tune was trained at a maximum of 16,384 tokens and engine.py says requests over the trained length limits are refused
- Inputs and outputs: A state (JSON or text) and named questions, each choice (1 to 255 options), noul (yes or no) or score (2 to 10 levels). All questions share one prefill of the state. It returns a probability for every option, the choice and a confidence. No text generation and no thinking
- Hardware: setup.sh says one CUDA GPU on the pytorch/pytorch:2.8.0-cuda12.8 image. The Decision Index ran its private tests on one RTX PRO 6000. No memory figure from Base Labs
- Training data: Described in the report, not published. Mix 8, 6.21M examples, 3.62M core decision skills, 1.00M safety and abstention, 1.00M judging and expert domains and 0.59M public decision data from 61 public datasets, generated with models such as DeepSeek V4-Pro and GLM-5.3 and filtered by a five-judge panel, cleaning, decontamination and review. No cut-off

- Scores: Capability pending, Reach 55, Risk 50, Agent interface 45, Documentation and upkeep 40 · negative events -2 · total over the 7 assessed categories
- Why:  · Reach, Memory 25 of 30, the BF16 repository is 18.8 GB, 22.6 GB with a fifth added, so it fits 24 GB. · Risk, File safety 15 of 25, safetensors only, but serving it means running the repository's setup.sh and patch. · Agent interface, Read as a decision model, by what it takes and returns. · Documentation and upkeep, Model card 5 of 20, two of the seven items.
- Sources: 22, open questions: 5, both in the full twin
- Capabilities: inference.decision, model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/elephant-1.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/elephant-1.svg` or a link to https://www.anchorterminal.com/tools/elephant-1 from a page on baseten.co or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Start on a CUDA GPU in pytorch/pytorch:2.8.0-cuda12.8-cudnn9-devel with `git clone https://huggingface.co/baselabs/elephant-1 && cd elephant-1 && bash setup.sh /tmp/w`, which clones OneJev at 81ce62f, applies serve.patch and installs the Decision Index kit
2. Then `qev serve --model . --port 8000` is OneJev's documented server. Base Labs doesn't document it for this checkpoint, so check the answers before relying on them
3. Don't load it with stock vLLM or Transformers. They ignore the block-bidirectional attention in config.json's modeling_overrides and give probabilities the model wasn't trained for
4. Keep each request under 16,384 tokens, the trained length, or the engine refuses it
5. Group the questions about one state in one request, since the state is read once and shared
6. Without hardware, Baseten's /v1/systemone serves it as decision-models/baseten-elephant-9b-latest with a Baseten key, free until the end of the launch week

## Connect

```bash
git clone https://huggingface.co/baselabs/elephant-1 && cd elephant-1
bash setup.sh /tmp/w   # clones OneJev at 81ce62f, applies serve.patch, installs the Decision Index kit (CUDA image pytorch/pytorch:2.8.0-cuda12.8-cudnn9-devel)
```

```bash
qev serve --model . --port 8000   # OneJev's System One server; its README documents the command, Base Labs doesn't
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| JEV-27B-VL | C | 59.5 | inference.decision, model.open-weights | https://www.anchorterminal.com/tools/jev-27b-vl.min.md |
| mammoth-1 | D | 47.7 | inference.decision, model.open-weights | https://www.anchorterminal.com/tools/mammoth-1.min.md |
| Qwen3-Embedding-8B | BB | 74.7 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| OpenAI Decisions API | BB | 73.1 | inference.decision | https://www.anchorterminal.com/tools/openai-decisions-api.min.md |


## Panel reviews (4, average 2.3/5, desk reviews from public material, no calls made)

- ★★☆☆☆ Safetensors weights served through a patched outside runtime (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Licence switched in a card edit on release day, with no note (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ A request format that exists only in engine.py (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★☆☆ 18.84 GB in BF16 fits one 24 GB card, with one runtime (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-22; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with ratings of 2 and 3. All four record that the licence moved from proprietary to MIT in the card's metadata on 21 October 2026 with no LICENSE file, and that serving it as trained needs OneJev at a pinned commit with the repository's patch. Read it as a 9B decision model that fits a 24 GB card in BF16, with a thin card and one patched runtime.