Vev 0.1.1

• The questions of a request run as one batch instead of one forward pass each. On one H800, 100 questions about a
  short text take 0.23 s instead of 3.75 s with vev-4b; images go through the image processor once per request.
• Requests with one question give bit-identical probabilities to 0.1.0. With several questions, probabilities move by
  up to a few hundredths (bf16 rounding); spec §6 and the conformance tolerance are updated with the measured values.
• flash-linear-attention 0.5.2 is a dependency on Linux, as in the evaluation environment; the server logs which
  DeltaNet kernels it uses. torchvision is a dependency.
• vev serve --revision; startup fails fast without CUDA; 500 responses no longer include the exception text.
• evals.run --one-question-per-request.

The weights are unchanged from 0.1.0 (collection): vev-4b · vev-4b-lora · vev-9b · vev-9b-lora.

Vev 0.1.0

First public release, a research preview.

• vev serve: HTTP server with the /v1/systemone request and response shapes. Text, JSON and images in state;
  noul, choice and score questions.
• Models: vev-4b and vev-9b, LoRA fine-tunes of Qwen3.5-4B and Qwen3.5-9B, released as merged weights and as
  adapters. Weights are CC BY-NC 4.0.
• Conformance suite that runs against any /v1/systemone implementation.
• Evaluation harness, dataset converters and the training code used for the release.

Weights and data on Hugging Face (collection):
vev-4b · vev-4b-lora · vev-9b · vev-9b-lora · vev-anchor-v3 (training anchor data).