mammoth

Checkpoints	Inputs
baselabs/mammoth-1, baselabs/elephant-1	text

mammoth-1 (27B) and elephant-1 (9B) are Base Labs' decision models on Qwen3.5. Their release serves them with OneJev's qev and a patch, so Flide serves them with the OneJev family: OneJev's prompt for every question, the answer slots read from the output head at the question's last token, and the calibration.json next to the weights when a checkpoint ships one (these two ship none).

They read a prompt in two blocks. In the full-attention layers the system turn and the state see each other in both directions, and so do the tokens of one question with its options and answer line; a question sees the state, and the state never sees a question. The Gated DeltaNet layers stay causal. The checkpoints say so in readout.json and under modeling_overrides in config.json, and Flide reads it at load. The state is still computed once and forked into the questions, each block read whole, and the state block ends where the question text starts, as in the release.

Speed

On one H200, one request at a time, Flide answers 1.6 to 4.9 times faster with mammoth-1 and 2.0 to 7.9 times faster with elephant-1 than their own engine.py on OneJev's qev with the checkpoints' serve.patch (transformers 5.19 with block masks, torch 2.8). Each workload is 32 requests, each run once on both sides before timing, and the table gives the median latency of a request. MultiRC requests ask about 49 questions each.

mammoth-1

Workload	Flide	Release	Speedup
Decision Index	61.9 ms	203.7 ms	3.3×
RACE	41.4 ms	201.0 ms	4.9×
MultiRC	433.9 ms	982.6 ms	2.3×
QuALITY long	352.5 ms	571.9 ms	1.6×

elephant-1

Workload	Flide	Release	Speedup
Decision Index	19.8 ms	104.7 ms	5.3×
RACE	12.9 ms	102.0 ms	7.9×
MultiRC	133.1 ms	447.5 ms	3.4×
QuALITY long	109.2 ms	219.5 ms	2.0×

Start a server

flide serve --model baselabs/mammoth-1

Send a request

curl http://localhost:8000/v1/systemone -H "Content-Type: application/json" -d '{
  "state": "Order #1182 arrived with a cracked screen. The customer asks for a refund and attached two photos.",
  "questions": {
    "action": {"type": "choice", "instructions": "What should support do next?",
               "criteria": {"refund": "refund the order", "replace": "send a replacement", "escalate": "ask a human"}},
    "urgent": {"type": "noul", "instructions": "Does this need a reply within the hour?"}
  }
}'

Limits and options

A request takes choice, noul and score questions, with OneJev's limits: a choice with more options than the model has answer slots is asked in chunks, so a choice takes up to 256 options, and a score takes as many levels as the model has slots. The longest state plus question is 32768 tokens and the longest request 65536. A request with images is refused, as the release reads text only.

Option	What it does
?debias=3	asks each question in 3 cyclic shifts of its options and combines them (default 1, the order given)