New AI Models

o4-mini

o4-mini is OpenAI's small reasoning model in the o-series, released on April 16, 2025 and built for fast, cost-effective reasoning. OpenAI describes it as the best-performing benchmarked model on AIME 2024 and AIME 2025, reaching 99.5% pass@1 on AIME 2025 when given access to a Python interpreter, and it outperforms its predecessor o3-mini on non-STEM tasks as well as domains like data science. It can combine web search, Python, and image analysis within a single chain of thought, and supports significantly higher usage limits than o3, making it a strong high-volume, high-throughput option for questions that benefit from thinking.

DolphinGemma

DolphinGemma is a roughly 400M parameter, audio-in audio-out AI model from Google, built on Gemma and trained on the Wild Dolphin Project's acoustic database of wild Atlantic spotted dolphins collected since 1985. It learns the structure of dolphin vocalizations such as signature whistles, burst-pulses, and click trains, and predicts likely subsequent sounds in a sequence, helping researchers surface recurring patterns in dolphin communication. Sized to run on Pixel phones in the field, it pairs with the CHAT system developed with Georgia Tech, which associates synthetic whistles with objects to explore shared communication. Google planned to share it as an open model, adaptable to other cetacean species.

GPT-4.1

GPT-4.1 is OpenAI's flagship API-only model released on April 14, 2025 alongside GPT-4.1 mini and GPT-4.1 nano. It outperforms GPT-4o in coding and instruction following, scoring 54.6% on SWE-bench Verified, 38.3% on Scale's MultiChallenge, and 87.4% on IFEval, while handling up to 1 million tokens of context and setting a 72.0% state-of-the-art result on the Video-MME long no-subtitles category. With a June 2024 knowledge cutoff, a 32,768-token output limit, lower cost, and faster latency than GPT-4o, it suits large codebase editing, long document analysis, and agent workflows built on the Responses API.

Gen-4 Turbo

Gen-4 Turbo is the faster, lower cost version of Runway's Gen-4 video generation model, exposed as gen4_turbo on the Runway API. It works as image-to-video: a required input image serves as the first frame and establishes subjects, style and lighting, while a text prompt of up to 1000 characters describes the motion, helping preserve character and scene consistency across shots. Clips run 5 or 10 seconds at 24 fps in six aspect ratios from 16:9 1280x720 to 21:9 1584x672. It costs 5 credits per second (25 credits for 5 seconds, 50 for 10) versus 12 for standard Gen-4, and Runway recommends it for quickly iterating ideas before switching to Gen-4 for final results.

SeedEdit 3.0

SeedEdit 3.0 is an instruction-based image editing model from ByteDance's Seed team, built on the Seedream 3.0 text-to-image model and served through the Volcano Engine Ark API under the model ID doubao-seededit-3.0-i2i. It applies edits described in natural language, including portrait retouching, background swaps, style changes, lighting adjustments, object removal and text rendering, while keeping untouched areas faithful to the original picture. ByteDance highlights that the model weighs which regions to change and which to preserve, and its in-house evaluation reports an image consistency score of 4.07 out of 5 and a usability rate of 56.1%, both ahead of SeedEdit 1.6. Typical uses include e-commerce imagery, portrait retouching and creative content production where a one-sentence edit should not degrade the rest of the photo.

kling-v2

kling-v2 is Kuaishou's Kling 2.0 video generation model from April 2025. It supports text-to-video and image-to-video with clear gains in motion quality, semantic responsiveness to prompts, and visual aesthetics, and it pairs with the multi-modal visual language (MVL) concept so creators can combine text with image and video references. In Kuaishou's internal GSB evaluations, the image-to-video mode reported a win-loss ratio of 182% against Google Veo 2 and 178% against Runway Gen-4. It fits short-form content, advertising, and cinematic concept work, available through the Kling web app and API.

Llama 3.1 Nemotron Ultra 253B v1

Llama 3.1 Nemotron Ultra 253B v1 is a 253B-parameter open reasoning model from NVIDIA, derived from Llama 3.1 405B Instruct via Neural Architecture Search based pruning and knowledge distillation. Reasoning can be toggled ON or OFF through the system prompt (detailed thinking on/off), and the BF16 model fits on a single 8xH100 node. The official model card reports Reasoning ON scores of 72.50 on AIME25, 66.31 on LiveCodeBench, 76.01 on GPQA and 97.00 on MATH500, with a 128K token context. Released under the NVIDIA Open Model License, it targets AI agent systems, chatbots, RAG pipelines and instruction-following applications.

PixVerse V4

PixVerse V4 is a video generation model release from AISphere's PixVerse platform. The base model was upgraded for more natural human motion, closer adherence to physics, and finer emotional expression, while new tools let creators add sound effects that match the on-screen action, generate lip-synced voice lines via Speech for both live-action and anime-style characters, and restyle existing footage with presets such as cyberpunk and felt or a custom prompt. The official announcement also cited fast-mode generation completing in as fast as 5 seconds, making V4 suited to short-form content, marketing clips, and character-driven videos that need synchronized audio and picture.

Wan2.1-KF2V-Plus

Wan2.1-KF2V-Plus is the keyframe-to-video model of Alibaba's Wan2.1 video generation family, served on Alibaba Cloud Model Studio. It takes a first-frame image, a last-frame image and a text prompt, then generates a smooth transition video between the two frames. Per the official documentation it outputs silent 5-second clips at 720P and 30fps in MP4 (H.264). The Plus tier focuses on quality: the release notes highlight complex large motions, physically plausible movement, rich artistic styles and cinematic detail with improved prompt following. Typical use cases include shot in-betweening, product showcases and storyboard previews where the start and end frames are fixed.

Llama 4 Maverick

Llama 4 Maverick is Meta's open-weight multimodal LLM released on April 5, 2025, built on a Mixture-of-Experts (MoE) architecture with early fusion for native multimodality. It has 17B activated parameters, 400B total parameters, 128 experts, and a 1M token context window. Meta's official model card reports instruction-tuned scores of 80.5 on MMLU Pro, 69.8 on GPQA Diamond, 43.4 on LiveCodeBench, and 73.4 on MMMU, ahead of Llama 3.1 405B on several benchmarks. It takes multilingual text and images (tested with up to 5 input images) and produces text and code across 12 officially supported languages. It fits assistant-style chat, visual reasoning, image captioning, long-document understanding, and synthetic data generation or distillation. Weights ship in BF16 and FP8, with FP8 fitting on a single H100 DGX host.

Llama 4 Scout

Llama 4 Scout is Meta's open-weight multimodal model released on April 5, 2025, alongside Llama 4 Maverick in the Llama 4 herd. Per the official model card, it is an auto-regressive language model built on a mixture-of-experts (MoE) architecture with 16 experts, 17B activated parameters and 109B total parameters, using early fusion for native multimodality that takes multilingual text and image input. Its context window reaches 10M tokens, the longest in the Llama 4 lineup, and Meta states it fits on a single 80GB H100 GPU with on-the-fly int4 quantization after pretraining on roughly 40 trillion multimodal tokens. Official evaluations report 69.4 on MMMU, 88.8 on ChartQA and 94.4 on DocVQA for the instruction-tuned model. It suits long-document analysis, image-grounded question answering, multilingual assistants and agentic workflows that need a large context window.

grok-3

grok-3 is xAI's flagship large language model, first previewed in February 2025 and opened to the xAI API in April 2025. Trained on the Colossus supercluster with 10x the compute of previous state-of-the-art models, it pairs extensive pretraining knowledge with strong reasoning across mathematics, coding, and world knowledge. Its Think mode, refined with large-scale reinforcement learning, lets the model reason for seconds to minutes, backtrack, and correct its own errors. Per the official announcement, grok-3 (Think) scored 93.3% on AIME 2025, 84.6% on GPQA, and 79.4% on LiveCodeBench, while the base model reached an Elo of 1402 in Chatbot Arena. It supports a 1 million token context window and works with DeepSearch, xAI's agent for real-time web research, fitting complex reasoning, STEM problem solving, code generation, long document analysis, and research tasks that need current information.

grok-3-mini

grok-3-mini is xAI's small reasoning model in the Grok 3 family, unveiled alongside Grok 3 in February 2025 and made available through the API on April 4, 2025. Trained with reinforcement learning to refine its chain-of-thought, it supports a thinking mode with adjustable effort (low or high), letting developers trade latency for deeper reasoning. xAI positions it as the cost-efficient option for STEM tasks that do not require extensive world knowledge: with high test-time compute it scored 95.8% on AIME 2024 and 80.4% on LiveCodeBench, and with reasoning off it still reached 66.2% on GPQA and 78.9% on MMLU-pro. It suits cost-sensitive applications that need fast math and code reasoning at scale.

V7

V7 is the new-generation image generation model that Midjourney released in alpha on April 4, 2025. The official announcement describes it as much smarter with text prompts, with noticeably higher image quality and beautiful textures, and significantly better coherence on hands, bodies and objects of all kinds. It is the first Midjourney model with personalization on by default, and it introduces draft mode, which renders at 10 times the speed for half the cost so ideas can be iterated quickly, then re-rendered at full quality with enhance or vary. V7 shipped in Turbo and Relax modes, with Turbo jobs costing 2x a normal V6 job, while upscaling, editing and retexture initially fell back to V6, with updates to follow. It suits concept exploration, style iteration and detail-critical image work.

Gen-4

Gen-4 is Runway's next-generation video generation model, released on March 31, 2025. Working mainly as image-to-video, it turns a single reference image plus a text prompt into 5 or 10 second clips without fine-tuning. Its core focus is consistency: characters, objects and locations stay stable across scenes, lighting and stylistic treatments. Gen-4 also adds shot-level camera control, realistic motion, stronger physics simulation and high prompt adherence, so results can sit beside live action, animated and VFX footage. It fits filmmaking, advertising and other storytelling work that needs repeatable, reference-consistent shots.

hunyuan-video-aries-fast

hunyuan-video-aries-fast is the identifier for the fast generation variant of Tencent Hunyuan's open source HunyuanVideo model, recorded as released on 2025-03-28. It belongs to the 13 billion parameter text-to-video family documented in the official HunyuanVideo repository, which uses a Dual-stream to Single-stream DiT, an MLLM text encoder, and a Causal 3D VAE to produce up to 720p videos of 129 frames (about 5 seconds). The official repository explicitly notes a released fast version alongside the high quality version used for benchmarks, trading some peak quality for lower latency and cost. The exact string is a codename without a public documentation entry, so exact specs of this variant should be confirmed with the serving platform. It suits quick video drafts, batch generation, and cost sensitive content pipelines.

hunyuan-video-i2v-aries-generalvideoseffect

hunyuan-video-i2v-aries-generalvideoseffect is the aggregator-style identifier for Tencent Hunyuan's image-to-video capability in its general video effect variant, where aries serves as an internal codename. Its foundation is HunyuanVideo-I2V, the image-to-video framework Tencent open-sourced in March 2025, which uses a token replace technique with a decoder-only MLLM text encoder to inject reference-image information into video generation, supporting output up to 720P and 129 frames (about 5 seconds). On top of plain image animation, the general video effect variant applies preset VFX templates such as one-shot camera moves, subtle subject motion, and hug or kiss interactions, turning a still photo into a stylized short clip. Typical use cases include short video creation, e-commerce showcases and social media content, and custom effects can be trained via LoRA and invoked with trigger words in the prompt.

3.0 (latest)

Ideogram 3.0 is a text-to-image model released by Ideogram on March 26, 2025, built around photorealism, accurate text rendering, and reusable style control. The official model page reports significant advancements in image-prompt alignment, photorealism, and text rendering quality, and states that human evaluations scored it highest in ELO rating among text-to-image models across diverse prompts. New Style References accept up to 3 reference images to steer the aesthetic of generations, with preferred looks reusable via Style Codes from a library of 4.3 billion style presets. It renders stylized, accurate text in complex and lengthy layouts and generates brand assets such as logos, posters, landing pages, and product photography, alongside natural skin tones, accurate reflections, and proper lighting physics. Typical use cases include graphic design, advertising, marketing, and campaign visuals.

3.0 (March 2025)

This entry is the dated snapshot of Ideogram 3.0, matching the model version released on March 26, 2025; it stays fixed rather than following the rolling 3.0 (latest) alias. Ideogram 3.0 is a text-to-image model with advances in image-prompt alignment, photorealism, and text rendering quality, and it scored the highest ELO rating in Ideogram's human evaluations across a diverse set of prompts. Features include Style References with up to 3 reference images, reusable Style Codes, a Random style feature drawing on a library of 4.3 billion presets, and Batch Generation for variants in a single session. It fits brand graphics, posters, landing pages, product photography, and photorealistic scenes that need accurate in-image text.

DeepSeek-V3 0324

DeepSeek-V3 0324 is the March 2025 upgrade snapshot of DeepSeek's open-source MoE model DeepSeek-V3 (671B total parameters, about 37B activated per token), served through the API as deepseek-chat. The official changelog reports clear reasoning gains over the previous snapshot: MMLU-Pro rose from 75.9 to 81.2, GPQA from 59.1 to 68.4, AIME from 39.6 to 59.4, and LiveCodeBench from 39.2 to 49.2. The update also improves front-end web development with more accurate code and better looking web pages and game front-ends, upgrades Chinese writing toward the R1 style with stronger medium-to-long-form quality, refines translation, letter writing, and multi-turn interactive rewriting, and ships more accurate function calling that fixes issues from earlier V3 versions. It suits chat assistants, coding helpers, and agent workflows that depend on reliable tool calling.