Converts multimodal inputs (text, image, audio, video) plus a question and options into calibrated probability distributions over choices. Built on Gemma 4 12B with a 30,000-question fine-tune, optimized for per-question decision classification and low-latency inference (Apache-2.0).
[next-mdx-remote-client] error compiling MDX: Unexpected character `=` (U+003D) before name, expected a character that can start a name, such as a letter, `$`, or `_` More information: https://mdxjs.com/docs/troubleshooting-mdx
Generates video with synchronized stereo audio from text, keyframes, or reference media in 4–8 sampling steps. Pruning, low-rank factorization, and INT8/GGUF formats reduce runtime costs, though speed and hardware needs vary by setup.
Cuts the FP4 checkpoint's weight size by about 16% by reducing routed experts from 512 to 384 per layer, while keeping 10 experts active per token. It also includes GRPO-style tuning in a complete merged checkpoint; CVSS arithmetic remains unreliable.
Provides parallel speculative-decoding drafts for Qwen3.8-27B. It passes predicted features between adjacent positions across layers, improving acceptance length and throughput in the reported SGLang benchmarks.
