LLM Models

Cosmos-Predict2.5-2B is a 2,059,174,912-parameter diffusion transformer in NVIDIA's Cosmos world foundation model family, aimed at physical AI such as robotics and autonomous vehicles. It unifies Text2World, Image2World and Video2World in one model and uses Cosmos-Reason1, a physical-AI reasoning VLM, as its text encoder. From a text prompt under 300 words plus an image or a 5-frame video it generates a 5-second 1280x704 mp4 at 16 FPS, requiring 32.54 GB of GPU VRAM. Weights are under the NVIDIA Open Model License; source code is Apache 2.0.

Universal-3.5 Pro is AssemblyAI's flagship pre-recorded speech-to-text model, requested as speech_models: ["universal-3-5-pro"] on /v2/transcript. It transcribes 18 languages with native mid-sentence code switching, takes contextual prompting plus keyterms prompting of up to 1,000 words, and solves transcription and speaker attribution jointly rather than with a separate diarizer. Requests in other languages fall back automatically to Universal-2, which covers 99. AssemblyAI's changelog dates the pre-recorded launch to July 7, 2026 and the separate streaming sibling to June 23, 2026.

Voxtral Mini Transcribe 2 (API id voxtral-mini-2602) is Mistral's batch speech-to-text model, described by the vendor as an efficient audio input model pre-trained and optimized for transcription purposes only. It is served from /v1/audio/transcriptions with speaker diarization, word-level timestamps and context biasing of up to 100 custom terms, accepts recordings up to 3 hours per request, and covers 13 languages. Mistral ships it in the Premier tier, so weights are not published, and prices it at $0.003 per minute.

Ink 2 is Cartesia's streaming speech-to-text model for real-time voice agents, exposed through the API as model id ink-2. It transcribes English only and has turn detection built in: rather than a separate VAD it emits a lifecycle of turn events (turn.start, turn.update, turn.eager_end, turn.resume, turn.end) driven by semantic endpointing that decides a turn is over by meaning rather than silence. Cartesia reports a time-to-final-transcript of 0.1s and accuracy on structured entities such as phone numbers, dates and email addresses.

Kyutai STT is a streaming speech-to-text model architecture from the Paris open-science lab Kyutai, released as two checkpoints: stt-1b-en_fr (~1B parameters, English and French, a 0.5 second delay, and a built-in semantic voice activity detector) and stt-2.6b-en (~2.6B parameters, English only, a 2.5 second delay). It is a decoder-only Transformer that consumes Mimi audio tokens at 12.5 Hz and emits a text stream time-aligned to the audio, producing punctuated transcripts with word-level timestamps. Weights are released under CC-BY 4.0.

MiniMax's open-weight omni-modal generation system. It takes a unified multimodal context of text, images, video and audio and produces video with native 32 kHz stereo audio at 24 fps, 4 to 15 seconds long, with the shorter side set to 768 pixels by default. The open release is H3-Base: two CFG-distilled checkpoints, FL2VA for text and first/last-frame input and Ref2VA for reference images, video and audio, built on a 33B-parameter dense single-stream Omni Transformer that draws hidden states from a Qwen3-VL-32B encoder. 2K output requires the hosted H3-Regenerate-2K module, which is not open sourced.

Lightricks' open-weight audio-video model, published on Hugging Face as a split, ComfyUI-aligned pack with one safetensors file per component rather than a single monolith. A 22B-parameter diffusion transformer paired with a custom Gemma 4 12B text encoder generates synchronized video and audio from text, image and video inputs, supported by separate video and audio VAEs, a vocoder, x2 spatial and temporal latent upscalers, and an optional duration head that infers clip length from the prompt. Both a distilled checkpoint on a fixed 8-step schedule at CFG 1 and a full trainable transformer are released.

Genmo's open text-to-video model, distributed as the mochi-1-preview checkpoint under Apache 2.0. It is built on a 10B-parameter Asymmetric Diffusion Transformer (AsymmDiT) whose visual stream carries nearly four times the parameters of its text stream, paired with a 362M-parameter AsymmVAE that compresses video 128x (8x8 spatial, 6x temporal) into a 12-channel latent space. Prompts are encoded by a single T5-XXL model. Genmo calls it a research preview that generates 480p video today, and publishes weights on Hugging Face, as a direct download and as a magnet link, alongside a free hosted playground.

Zhipu AI's open-weight text-to-video model, released 8 November 2024 as the newest generation of the CogVideoX family. It generates 1360x768 video at 16 frames per second in 5- or 10-second lengths, with frame counts of 16N+1 where N is at most 10 (default 81). Prompts must be English and are capped at 224 tokens. The vendor reports diffusers inference from 10 GB of VRAM in BF16 and from 7 GB with INT8, against 76 GB for the SAT BF16 path, and roughly 1000 seconds on a single A100 for a five-second clip. Image-to-video lives in a separate checkpoint, CogVideoX1.5-5B-I2V.

Tencent's open-weight video generation model, released with inference code and weights on 20 November 2025. An 8.3B-parameter Diffusion Transformer paired with a 3D causal VAE (16x spatial and 4x temporal compression) handles both text-to-video and image-to-video at 480p and 720p, and a separate few-step super-resolution network upscales output to 1080p. The default generation length is 121 frames, exported at 24 fps. Tencent states a minimum of 14 GB of GPU memory with model offloading enabled, and a step-distilled 480p image-to-video checkpoint that generates on a single RTX 4090 within 75 seconds.

The largest reranker in Alibaba's Qwen3 Embedding series: an 8B-parameter, 36-layer cross-encoder built on the Qwen3 foundation model. It scores a query-document pair by reading the model's yes/no token logits under a judging prompt, and is instruction-aware, so a task description tailored to a scenario or language can be supplied per request. It covers over 100 natural and programming languages and is normally paired with a Qwen3-Embedding model as the first-stage retriever. Weights ship on Hugging Face in Safetensors under Apache 2.0 and run through the sentence-transformers CrossEncoder or Transformers; 0.6B and 4B rerankers complete the series.

The largest text embedding model in Alibaba's Qwen3 Embedding series, an 8B-parameter, 36-layer dual-encoder built on Qwen3-8B-Base and fine-tuned with LoRA for embedding and ranking work. It emits vectors of up to 4096 dimensions with user-defined output dimensions from 32 to 4096 (Matryoshka), is instruction-aware so a task description can be prepended to queries, and covers over 100 natural and programming languages for multilingual, cross-lingual and code retrieval. Weights ship on Hugging Face in Safetensors under Apache 2.0 and run through sentence-transformers, Transformers, vLLM or Text Embeddings Inference; 0.6B and 4B siblings cover lower-resource deployments.