Baseten open-sources elephant-1, a 9B decision model at a quarter the cost
Balance- · reddit · 2026-10-24
Baseten's Base Labs released open weights for elephant-1, a 9B MIT-licensed decision model post-trained from Qwen3.5-9B. It doesn't generate text: given a state and yes/no or multiple-choice questions, it returns probabilities read from option-letter logits, with all questions sharing one prefill.
- Artificial Analysis scored it 65 on the Decision Index, tying d1-ultra at $8 per 1M tasks vs $31; best under 15B on the Jev Index (62.65). Its 27B sibling mammoth-1 ranks #1 at 71.24.
- Sweet spot: routing, guardrails, tool-call gating, LLM-as-judge. It deliberately says "I don't know" a lot (93 non-hallucination, 12 accuracy on Omniscience).
- Full-attention layers use block-bidirectional masking for prefix-cache reuse (48.6 vs 44.8 causal). Training: 6.21M examples, 5-judge labeling, decontamination; scaling data 16x took Jev from 45.5 to 66.0, source auditing added 8 more.
- Loss is cross-entropy + Brier with a trained "I don't know" option; 1 epoch on 128 H100s. Self-hosts on one CUDA GPU; free API until Oct 27.
Related event: Baseten Open-Sources Decision Models mammoth-1 and elephant-1(7 posts)→
More from coding & agent
- GPT SOL 6.1 builds a 3D axe from scratch via Blender MCP in 7 minutes — OpenAIDevs · 2026-10-24
- icons.computer indexes 3,000+ curated app icons, agent-ready via Firecrawl — devdigest · 2026-10-24
- Dev: With LLMs Helping Setup, There's Never Been an Easier Time to Daily Drive Linux — KevinNaughtonJr · 2026-10-24
- Postman's first AI-Ready API leaderboard puts You.com (96) above OpenRouter (94) — RichardSocher · 2026-10-24
- Richard Socher launches Flint, an open-weights local coding model hitting 125+ tokens/s on a MacBook — RichardSocher · 2026-10-24
- Agent platform Outpost teases mobile app to manage all your agents from one place — jacob_posel · 2026-10-24