LLM Models

Vidu Q3 is the third generation of the Vidu video model family from Shengshu, offered through the Vidu platform API as viduq3-pro, viduq3-mix and viduq3-turbo, plus the industry-tuned viduq3-drama and viduq3-ad. The core variants render 1 to 16 second clips at 540p, 720p or 1080p at 24fps and support simultaneous audio and video output, with image-to-video, text-to-video and reference-to-video paths.

Gen-4.5 is Runway's flagship video generation model, callable through the Runway Dev API as model id gen4.5 and available in the Runway app on all paid plans. It accepts text or image input and returns video, and in Tool Mode and Edit Studio its generations can be exported as ProRes 4444 with PCM audio or as a PNG sequence rather than a compressed MP4. It is billed at 12 credits per second of output.

MiniMax-Hailuo-2.3 is a proprietary video generation model exposed through the MiniMax platform API, callable on both the text-to-video and image-to-video endpoints as model value MiniMax-Hailuo-2.3. It generates 6- or 10-second clips at 768P, or 6-second clips at 1080P, from a prompt of up to 2000 characters that may include camera-movement commands. A cheaper MiniMax-Hailuo-2.3-Fast variant shares the same endpoints.

Stable Video Diffusion (SVD) Image-to-Video XT is Stability AI's latent diffusion model that takes a still image as a conditioning frame and generates a short video from it. The XT checkpoint is finetuned to produce 25 frames at 576x1024 resolution, using a finetuned f8-decoder for temporal consistency; a frame-wise decoder is also provided. The model cannot be controlled through text and produces clips of about four seconds or less.

BAAI's Apache-licensed multilingual reranker, a 0.6B cross-encoder built on the bge-m3 base. Unlike an embedding model it takes a query and a passage together and outputs a relevance score directly, which can be mapped to [0,1] with a sigmoid. It is the lightweight recommendation in BAAI's own reranker lineup and pairs naturally with the bge-m3 embeddings already in this catalog.

Resemble AI's MIT-licensed open-source text-to-speech family, offering zero-shot voice cloning from a short reference clip. It ships as Turbo (350M, English, sub-200ms latency), Nano (110M, English, roughly 3x faster than realtime on 8 CPU cores) and Multilingual V3 (500M, 23+ languages), plus dedicated single-language finetunes. Every generated file carries Resemble's Perth perceptual watermark, which survives MP3 compression and common editing.

An Apache-licensed text-to-speech model at just 82 million parameters, covering 8 languages and 54 voices. Built on StyleTTS 2 with an ISTFTNet vocoder and a decoder-only design - no diffusion and no encoder release - it runs comfortably on CPU while delivering quality its maintainers describe as comparable to far larger models. It was trained on a few hundred hours of permissive, non-copyrighted audio for roughly $1,000 in total compute.

Alibaba's Apache-2.0 mixture-of-experts model, released April 2026, with 35B total parameters and only 3B activated per token. It accepts image, text and video input and emits text, natively handles a 262,144-token context (extensible to roughly 1.01M with YaRN scaling), and adds a thinking-preservation feature that carries reasoning context across conversation turns.

Google DeepMind's image generation and editing model, formally named Gemini 3.1 Flash Image and marketed as Nano Banana 2. It pairs Nano Banana Pro's capability with Flash-level speed, holds character consistency across up to 5 characters and 14 objects, renders legible text directly into images, grounds generations in real-world knowledge via web search, and upscales to 2K and 4K.

Reve AI's layout-first text-to-image model, released July 2026 as the successor to Reve 2.0. It renders at a full 16 megapixels and improves layout planning, spatial reasoning and foreign-text rendering, with every element in the generated image individually addressable and editable rather than baked into pixels.

Zhipu's 753B-parameter flagship, built for long-horizon agentic work and released under the MIT licence with no regional restrictions. It sustains a 1M-token context - a five-fold jump over GLM-5.1's 200K - using an IndexShare architecture that reuses one indexer across every four sparse-attention layers, cutting per-token FLOPs by 2.9x at full context length. A reasoning_effort parameter selects between 'high' and 'max' thinking budgets.

Mistral's document-intelligence model, released June 2026. It parses PDF, DOC, PPT and OpenDocument files across 170 languages in 10 language groups, handling interleaved imagery, mathematical expressions, tables and complex layouts. Every request returns markdown-structured text alongside bounding boxes, block types and confidence scores.