New AI Models

Wan2.1-I2V-Plus

Wan2.1-I2V-Plus is an image-to-video model from Alibaba's Tongyi Wanxiang (Wan) 2.1 series, served through the Alibaba Cloud Model Studio API. It takes a first-frame image plus a text prompt and generates a silent video at a fixed output of 720P, 5 seconds, 30fps in MP4 (H.264) format. Positioned as the quality tier of the Wan2.1 I2V lineup, alongside the faster Turbo variant that offers 480P/720P and 3 to 5 second options, it fits short-video creation, marketing assets, and e-commerce content where visual fidelity and motion consistency matter.

Finetuning

Finetuning is Black Forest Labs' customization service for FLUX.1 [pro]: upload a small dataset of images and train a private finetuned model that captures a specific subject, character, or style while keeping the base model's versatility. Training works with as few as 1 to 5 example images, and the resulting finetune runs across the whole FLUX model suite, including FLUX1.1 [pro] and FLUX Tools, at resolutions up to 4 megapixels without additional adaptation. In BFL's user study, its results were preferred 68.9% of the time over other finetuning services built on FLUX.1 [dev]. It also combines with FLUX.1 Fill [pro] for personalized inpainting and FLUX.1 Depth [pro] for structure-controlled generation, fitting use cases such as brand consistency, character consistency, marketing assets, and storytelling.

DeepSeek-V3.1

DeepSeek-V3.1 is a hybrid inference model from DeepSeek built on a single set of weights that supports both a thinking mode for deep reasoning and a non-thinking mode for direct answers, switched through the chat template. After post-training for tool use, it delivers major improvements on agent tasks and coding, scoring 66.0 on SWE-bench Verified and 31.3 on Terminal-bench, while its thinking mode reaches answers in less time than DeepSeek-R1-0528. The model serves both modes through the deepseek-chat and deepseek-reasoner endpoints with 128K context, making it suited for agent workflows, code generation, and long-context reasoning.

Wan2.1-T2V-Plus

Wan2.1-T2V-Plus is the quality-first text-to-video tier of Alibaba Cloud's Wan2.1 family, offered on Model Studio. Given a text prompt, it generates a silent 5-second 720P video at 30fps in MP4 (H.264). Compared with the faster Turbo variant, this Plus tier emphasizes visual quality and stability, making it a fit for final renders in marketing clips, short creative videos, and other prompt-driven production scenarios.

Wan2.1-T2V-Turbo

Wan2.1-T2V-Turbo is the lower-priced Turbo tier of Alibaba's Wan2.1 text-to-video family, served on Alibaba Cloud Model Studio as wan2.1-t2v-turbo (Singapore) and wanx2.1-t2v-turbo (Beijing). From a text prompt it generates 5-second silent MP4 clips at 30 fps in 480P or 720P, while the quality-focused wan2.1-t2v-plus tier runs only at 720P. In the Beijing region it is billed at 0.24 CNY per second of output versus 0.70 for plus, making it a practical option for social clips, bulk drafts and other high-volume workflows where cost matters more than peak fidelity.

DeepSeek-V3

DeepSeek-V3 is DeepSeek's open-source flagship Mixture-of-Experts (MoE) language model released in December 2024, with 671B total parameters and 37B activated per token, pretrained on 14.8T tokens. Its architecture combines Multi-head Latent Attention (MLA), DeepSeekMoE, an auxiliary-loss-free load balancing strategy, Multi-Token Prediction (MTP) and FP8 mixed precision training, and the official repo reports about 2.788M H800 GPU hours for full training. On official benchmarks the chat model scores 88.5 on MMLU, 90.2 on MATH-500 and 39.2 on AIME 2024, on par with GPT-4o and Claude 3.5 Sonnet, with a 128K context window. Served through the deepseek-chat API and released with MIT-licensed code plus commercially usable weights, it fits general chat, coding, math reasoning and cost-sensitive large-scale deployments.

Imagen 3

Imagen 3 is Google's third-generation text-to-image model. The December 2024 update brought it to ImageFX in more than 100 countries. Google says it generates brighter, better composed images, renders a wider range of art styles with greater accuracy from photorealism to anime, follows prompts more faithfully, and produces richer details and textures. In side-by-side comparisons judged by human raters, its outputs achieved state-of-the-art results against leading image generation models. It also powers the Whisk experiment in Google Labs and serves enterprise creative workflows through Vertex AI.

Veo 2

Veo 2 is Google DeepMind's text-to-video model announced in December 2024 as the successor to the original Veo. It generates high-quality videos across a wide range of subjects and styles at resolutions up to 4K and lengths extended to minutes, with a stronger grasp of real-world physics and the nuances of human movement and expression. It follows cinematography cues in the prompt, such as lens choice, shallow depth of field, and shot types like low-angle tracking shots and close-ups, and produces fewer artifacts such as extra fingers or unexpected objects. Google reports state-of-the-art results against leading models in head-to-head comparisons judged by human raters. The model rolls out gradually via VideoFX, YouTube and Vertex AI, all outputs carry an invisible SynthID watermark, and wider product availability is set to follow through 2025.

DeepSeek VL2

DeepSeek-VL2 is the standard variant in DeepSeek's Mixture-of-Experts (MoE) vision-language model series, activating about 4.5B of 27.5B total parameters per token. Built on DeepSeekMoE with Multi-head Latent Attention (MLA) and a dynamic tiling vision encoding strategy for high-resolution images, it covers visual question answering, OCR, document, table and chart understanding, visual grounding, and interleaved image-text dialogue within a 4096 token context, fitting document parsing, chart analysis, and multimodal assistant scenarios.

DeepSeek VL2 Small

DeepSeek-VL2-Small is the mid-tier variant of the DeepSeek-VL2 family, a series of Mixture-of-Experts (MoE) vision-language models released by DeepSeek on 2024-12-13. It combines a 16.1B-parameter DeepSeekMoE language model with a vision encoder, activating only about 2.4B parameters per token, and supports a 4096-token context. Per the official repository, the model covers visual question answering, OCR, document, table and chart understanding, and visual grounding, emitting bounding boxes for referenced objects through special tokens, and it accepts multi-image and interleaved image-text input; the README reports competitive or state-of-the-art results among open-source models with similar or fewer activated parameters. Standard inference may require an 80GB GPU, while incremental prefilling with chunk size 512 keeps it within 40GB. The code is MIT licensed and the model permits commercial use under the DeepSeek Model License, making it a fit for document parsing, image QA, and grounding-based applications.

DeepSeek VL2 Tiny

DeepSeek-VL2-Tiny is the smallest member of the DeepSeek-VL2 series, a family of Mixture-of-Experts (MoE) vision-language models from DeepSeek. It carries 3.37B total MoE parameters with only 1.0B activated per token and runs on a single GPU under 40GB of memory. The model combines a dynamic tiling vision encoding strategy for high-resolution images of varied aspect ratios with a DeepSeekMoE language model that uses Multi-head Latent Attention to compress the KV cache for efficient inference. Within a 4096-token context it handles visual question answering, OCR, document, table and chart understanding, and visual grounding that outputs bounding boxes via special tokens. Released on 2024-12-13 under the DeepSeek Model License with commercial use supported, it fits cost-sensitive multimodal workloads such as document parsing, image-based QA and image-text retrieval.

Phi 4

Phi 4 is Microsoft's 14B parameter small language model, released on December 12, 2024 under the MIT license and available on Azure AI Foundry and Hugging Face. It is a dense, decoder-only Transformer trained on 9.8T tokens with a 16K context window, using heavy volumes of textbook-like synthetic data plus curated organic and academic data. It specializes in complex reasoning, especially math: in Microsoft's official SimpleEval comparison it scores 56.1 on GPQA and 80.4 on MATH, above GPT-4o (50.6 and 74.6) and larger models such as Llama-3.3 70B, and outperforms Gemini Pro 1.5 on math competition problems. It fits research, reasoning-heavy workloads, latency bound scenarios and memory or compute constrained deployments, primarily for English text.

Gemini 2.0

Gemini 2.0 is the second generation of Google's Gemini model family, announced on December 11, 2024 with Gemini 2.0 Flash Experimental as the first release. It is built for the agentic era: the model can natively call tools such as Google Search, code execution, and user-defined functions, plan multi-step tasks, and follow complex instructions. It is natively multimodal, accepting image, video, and audio input and generating mixed text-and-image output plus steerable multilingual text-to-speech audio. Google positions 2.0 Flash as a low-latency workhorse that outperforms 1.5 Pro on key benchmarks at twice the speed, and pairs the family with agent prototypes such as Project Astra, Project Mariner, and Jules. It suits developers building agents, real-time multimodal applications, and in-product assistants, with general availability and more model sizes to follow.

Llama 3.3 70B Instruct

Llama 3.3 70B Instruct is Meta's open instruction-tuned text model released on December 6, 2024. It uses an optimized transformer architecture with GQA, offers a 128K token context window, and officially supports 8 languages (English, German, French, Italian, Portuguese, Hindi, Spanish, Thai) after pretraining on 15T+ tokens. According to Meta's model card, this 70B model reaches Llama 3.1 405B level results on many benchmarks: IFEval 92.1 vs 88.6, MATH 77.0 vs 73.8, GPQA Diamond 50.5 vs 49.0, plus HumanEval 88.4, MMLU 86.0 and MGSM 91.1. It targets assistant-like chat, multilingual text generation and coding, and is widely deployed for commercial and research use under the Llama 3.3 Community License.

GenCast

GenCast is Google DeepMind's AI ensemble weather forecasting model, presented in Nature in December 2024. It is a diffusion model adapted to the spherical geometry of the Earth, trained on four decades of historical weather data from the ECMWF ERA5 archive at 0.25 degree resolution. Instead of a single deterministic estimate, it generates 50 or more possible weather trajectories up to 15 days ahead. Tested on 2019 data across 1320 combinations of weather variables and lead times, it outperformed ECMWF's operational ENS on 97.2% of targets, and on 99.8% at lead times beyond 36 hours. A single Google Cloud TPU v5 produces one 15-day ensemble member in 8 minutes, with all members running in parallel. GenCast consistently outperformed ENS on extreme heat, cold, and high wind forecasts and on tropical cyclone track prediction, and was more accurate in a wind power generation experiment. Its code and weights are open, with real-time and historical forecasts to follow. It suits extreme weather risk preparedness, renewable energy planning, and disaster response.

Genie 2

Genie 2 is a large-scale foundation world model from Google DeepMind that turns a single prompt image into an action-controllable, playable 3D interactive environment, navigated with keyboard and mouse. Built as an autoregressive latent diffusion model trained on large video datasets, it can sustain consistent worlds for up to a minute, remembers elements that leave the current view, simulates effects such as water, smoke, gravity and reflections, and models NPC behavior and character animation across first-person, third-person and isometric perspectives. It is designed for training and evaluating embodied agents in a wide variety of novel environments, and also supports rapid prototyping by turning concept art into playable environments.

hunyuan-image-expansions

hunyuan-image-expansions is the image expansion (outpainting) variant of Tencent Hunyuan's image generation service, offered through the ImageOutpainting API on Tencent Cloud. It extends an input image beyond its original borders to a specified aspect ratio such as 1:1, 4:3, 3:4, 16:9, or 9:16, generating new content around the existing picture, and the target ratio must differ from the original one. The API accepts images with sides under 5000px in common formats including jpg, png, and webp, and returns the result as base64 or a temporary URL. It fits workflows that need to reframe or enlarge compositions, such as e-commerce visuals, social media posts, and advertising layouts.

hunyuan-video-aries

HunyuanVideo is Tencent's open-source text-to-video foundation model with over 13 billion parameters, the largest open-source video generation model at its release. It combines a Diffusion Transformer (DiT) with a dual-stream to single-stream hybrid design, a pretrained MLLM text encoder, and a Causal 3D VAE, generating 720p videos of up to 129 frames across multiple aspect ratios, with a prompt rewrite module for better instruction following. In the official evaluation over 1,533 prompts by more than 60 professional reviewers, it ranked first overall against closed-source baselines including Runway Gen-3 and Luma 1.6. Its open weights make it a practical base for creative video production, advertising, short-form content, and video generation research.

Nova Lite

Amazon Nova Lite is a very low-cost multimodal understanding model in the Amazon Nova family, announced at AWS re:Invent 2024 and available exclusively in Amazon Bedrock. It is built for fast processing of image, video, and text inputs to generate text output, accepting up to 300K tokens of input and analyzing multiple images or up to 30 minutes of video in a single request. AWS positions it for high-volume, price-sensitive workloads such as real-time customer interactions, document analysis, and visual question answering with high accuracy. It also supports text and multimodal fine-tuning, plus model distillation, to tune quality and cost for a specific use case.

Nova Micro

Nova Micro is the text-only model in Amazon's Nova family, available on Amazon Bedrock. According to the AWS announcement, it delivers the lowest latency responses in the Nova family at a very low cost, with a 128K token context window optimized for speed and cost. It fits text summarization, translation, content classification, interactive chat, brainstorming, and simple mathematical reasoning and coding, and supports customization on proprietary data through fine-tuning and model distillation. Unlike the multimodal Nova Lite, Pro, and Premier, Micro targets high-volume lightweight text tasks where response speed and price matter most.