LLM Models

SUPERSEDED. Zhipu AI's flagship multilingual model released January 2024 (API) with open-source GLM-4-9B variants in June 2024. Pretrained on 10 trillion tokens primarily Chinese/English plus 24 additional languages. Features 128K context length for both GLM-4 and GLM-4 All Tools. Advanced capabilities: web browsing, code execution, custom tool calls (Function Call), long context reasoning. Approaches state-of-the-art models (GPT-4 Turbo, Gemini 1.5 Pro, Claude 3 Opus) on standard benchmarks. Best-in-class for Chinese language tasks. GLM-4-9B open variant enables research and deployment on consumer hardware. Zhipu no longer documents this January 2024 model; the repository this record cited now hosts the GLM-4-0414 open-weight successor series instead. Unsourceable fields have been removed.

SUPERSEDED. 01.AI's proprietary flagship model with 70B parameters, released May 2024. Chinese AI powerhouse trained from scratch with significant improvements over Yi-34B open-source model. Performs on par with GPT-4 and Claude 3 across benchmarks. Exceptional multilingual capabilities - excels in Spanish, Chinese, Japanese, German, and French per LMSYS multilingual leaderboard. Positioned first on AlpacaEval 2.0 shortly after launch. Trained on 3.1T tokens of English and Chinese using cascaded deduplication and quality filtering. Uses decoder-only transformer with pre-normalization, SwiGLU, RoPE, and Grouped Query Attention. 01.AI no longer documents this model; its published API line-up is led by yi-lightning. The fields that could not be read at a vendor URL have been removed rather than left asserted.

DISCONTINUED. OpenAI's original text-to-video diffusion model, announced February 2024 (public release December 2024 as 'Sora Turbo'). Generated videos up to one minute long using a transformer architecture that represents video as patches. Superseded by Sora 2; the entire Sora product is being discontinued (app/web ended April 26, 2026; API ends September 24, 2026).

ByteDance's video generation model (2026), powering the Dreamina/Doubao platforms. Seedance 2.0 accepts quad-modal inputs (text, image, video, audio) and features a 'Universal Reference' system that locks composition, camera movement, and character actions across shots for consistent multi-shot sequences.

Tencent's open-weight text-to-video model (December 2024), a 13B-parameter model released with public weights that brought open video generation close to closed-source quality at launch. It produces high-motion, coherent clips and anchors a growing open video ecosystem.

Alibaba's open-weight cinematic video model (2025), supporting text-to-video and image-to-video at 720p and 1080p with various aspect ratios. Wan 2.x weights are downloadable from Hugging Face under Apache 2.0, making it a leading open option for self-hosted video generation.

MiniMax's Hailuo video model, known for fast render speed and surprisingly high quality, including a capable free tier. It delivers natural motion and decent prompt adherence for short-form content, making it popular for rapid ideation and experimental clips.

Luma AI's current-generation video model (Ray3), regarded as one of the best image-to-video tools in 2026 and strong for cinematic mood and environment-heavy shots. It supersedes the earlier Dream Machine naming and targets high-fidelity, physically coherent motion.

Runway's professional video generation model (March 2025), the pro favorite for granular creative control. Gen-4 emphasizes reference-driven character and world consistency, camera moves, and motion brush, making it a production pick for teams that need editing control alongside model quality.

Kuaishou's flagship AI video model, consistently ranked among the best overall for cinematic quality in 2026. Kling v3 improved temporal consistency so objects and people keep their shape across a clip, generating up to 10 seconds at 1080p with complex camera movement, from text and image prompts.

SUPERSEDED. Pika Labs' AI video generation platform enabling text-to-video, image-to-video, and video-to-video transformations. Released November 2023 with $55M funding. Generates 3-second clips in seconds to tens of seconds, outputs up to 1080p. Supports diverse styles: 3D animation, anime, cartoon, cinematic. Features: custom aspect ratios (square, portrait, widescreen for Instagram/TikTok), camera controls (pan, zoom, rotate), AI upscaling, video outpainting/inpainting for seamless editing. Web and Discord platform access. Free downloads, no watermark. Represents major milestone in accessible AI video generation. Pika no longer documents version 1.0; its current offering is Pika 2.5. The seeded pricing claim was the inverse of the vendor's plan table and has been removed rather than replaced, because the vendor's published prices describe Pika 2.5 and not this version.

Ideogram's groundbreaking text-to-image model with exceptional text rendering accuracy. Released August 2024, trained from scratch to significantly outperform competitors on image-text alignment, subjective preference, and text rendering. Achieves 85-90% text rendering accuracy vs competitors producing garbled text (DALL-E, Midjourney). Five style options: regular, realistic (photo-quality), design (premium graphics), 3D, anime. Features color palette control for brand consistency. Human evaluations consistently rate it as significant improvement over Flux Pro and DALL-E 3. iOS app and API (beta) available. API pricing lower than DALL-E 3, on par with Flux Pro.