Base Labs Released open-source decision models,mammoth-1 tops the Jev rankings
BIBIBI
AT A GLANCE
Under Baseten,Base Labs released open-source decision models and claimed leading results in evaluations.
Article
Baseten 1 Qwen transformed in weeks: 27B decision models top the rankings, 9B surpassing Jev
Beating AI Flash,AI the research team under inference infrastructure company Baseten Base Labs released open-source decision models mammoth-1 and elephant-1, with parameter counts of 27000000000 and 9000000000.
The team used only 1 weeks to complete post-training based on Qwen3.5-27B and Qwen3.5-9B, using 6210000 training samples.
On the Jev Decision Index 0.4 leaderboard,mammoth-1 ranked first with 71.24 points, ahead of another 207 decision models.elephant-1 scored 62.65 points, surpassing Jev 1.14's 62.2 points, becoming the highest-scoring model below 15000000000 parameters.
In another evaluation by Artificial Analysis,elephant-1 also tied for first with Liquid AI's d1-ultra.
The weights of both models have been made available under the MIT license and support self-deployment.
Original link https://m.theblockbeats.info/flash/373191
Key points
01
Beating AI Flash reports that the research team under Baseten Base Labs released the open-source decision models mammoth-1 and elephant-1。
02
mammoth-1 and elephant-1 have parameter counts of 27000000000 and 9000000000。
03
The team said it used only 1 weeks, based on Qwen3.5-27B and Qwen3.5-9B, to complete post-training, using 6210000 training samples.
04
The original article states that on the Jev Decision Index 0.4 leaderboard,mammoth-1 ranked first with 71.24 points, ahead of another 207 decision models.
05
The original article states that elephant-1 scored 62.65 points, surpassing Jev 1.14's 62.2 points, becoming the highest-scoring model below 15000000000 parameters.
06
The original article states that in Artificial Analysis's evaluation,elephant-1 tied for first with Liquid AI's d1-ultra.
07
The weights of both models have been made available under the MIT license and support self-deployment.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
Base Labs Post-trained existing base models, producing mammoth-1 and elephant-1. The report's focus is that they have different parameter scales, but both achieved high rankings in the decision-model evaluations mentioned in the article, and their model weights can be obtained and self-deployed under the MIT license.
Why it matters to readers
If the relevant leaderboard results can be supported by the original evaluations and independent replication, this suggests that relatively short post-training may also produce decision models with strong evaluation results that can be self-deployed.
Beginners can understand “parameters” as an indicator of model size and “post-training” as continued training on top of a base model; leaderboard scores reflect only the corresponding tests and cannot by themselves demonstrate capability across all real-world scenarios.
Risks and unknowns
The material does not provide Base Labs or Baseten's official release announcement or model-weight address.
The material does not provide the original evaluation pages, test methodology, or complete results for the Jev Decision Index and Artificial Analysis.
1 The claimed completion of post-training in 6210000 weeks, the training samples, and the various rankings cannot yet be independently verified based on the supplied material.
The models' actual deployment performance, community replication status, and MIT license compliance status are unknown.
Related Developments
Loading event timeline…
Related concepts
parameters
This term is not in the glossary yet. Browse related concepts in the glossary.
