Generates synchronized audiovisual output from text, image, or audio prompts — a diffusion-based multimodal model with componentized weights (video/audio VAEs, multilingual text encoder, distilled transformer) and ready integration with HuggingFace pipelines and ComfyUI.
Experimental MiniMax H3 variant that injects learned stylistic and motion 'character' from LTX 2.3, Wan 2.2 and Krea 2 into H3 by surgically grafting attention and MLP components; preserves H3 modality routing while shifting t2v/i2v aesthetics, with limited audio impact and community-license constraints.
A 9B dense reasoning LLM optimized for single‑GPU deployment and terminal-based coding agents, with long-context support (up to 262,144 tokens) and GGUF/quantized builds for edge/mobile. Strong on coding and agentic benchmarks.
Introduces Declarative Attention (DA), a zero-shot protocol that has LMs declare which parts of long context to attend to during chain-of-thought, letting the runtime build dynamic attention masks and skip most KV-cache reads. Produces large token savings (up to ~52% on Gemma-4-31B) with modest accuracy loss.
Evaluates whether AI agents can autonomously use Sparse Autoencoders (SAEs) to discover interpretable features and causally steer model behavior. Introduces a 20-task benchmark that has agents design contrastive probes to search a 131K+ feature dictionary in Gemma-2-9B-IT and evaluates activation rank, selectivity, and causal steering against expert references.
Rewrites AI-generated English and Chinese drafts so they read like human writing while preserving every number, date, unit, name and quote. Runs locally with multiple GGUF quantized builds and a strict byte-for-byte prompt format for consistent rewrites.
Generates unified 768‑dimensional embeddings for text (including code), images, video and audio to enable cross‑modal semantic search and retrieval. Supports task instruction prefixes, Matryoshka truncation to 128/256/512/768 dims, and modular encoders for on‑device use under an Apache‑2.0 license.
Analyzes why on-policy distillation can produce excessively long student rollouts by identifying termination-token mismatch between students and teachers, and proposes aggregating functionally equivalent EOS tokens as a shared semantic stopping action to mitigate length inflation.
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).
Provides calibrated probabilistic decisions (yes/no, 2–256 choice, 0–5 score) in one forward pass, with an optional adaptive-thinking mode that invokes Gemma‑4 when System 1 is uncertain; supports text+image, 256K context and vLLM serving, but adaptive thinking is much slower.
Returns calibrated probabilities for yes/no and multi-choice decisions using a two-stage System 1 (fast classifier) and System 2 (Gemma‑4 reasoning) pipeline; this NVFP4 release quantizes the 3,840 routed MoE experts to 4-bit so the model fits ≈17–18 GB of GPU memory while other weights remain bf16.
GGUF-packaged weights for EmbeddingGemma 2 enabling local multimodal embeddings (text, image, video, audio); supports 768/512/256/128 dimensions, BF16/FP32 inference, selective modality loading for lower memory, and is suited for semantic search and RAG.