New AI Models

Gemini 2.5

Gemini 2.5 is Google's family of thinking models, first released on March 25, 2025 with Gemini 2.5 Pro Experimental. Built on an enhanced base model with improved post-training, it reasons through its thoughts before responding. At launch it debuted at number one on the LMArena leaderboard by a significant margin, scored 63.8% on SWE-bench Verified with a custom agent setup, and reached 18.8% on Humanity's Last Exam without tool use, while also leading GPQA and AIME 2025. The models are natively multimodal across text, audio, images, video and full code repositories, with a 1 million token context window (2 million to follow). Suited to complex coding, math and science workloads as well as long-context analysis.

Llama 3.1 Nemotron Nano 8B V1

Llama 3.1 Nemotron Nano 8B V1 is an 8B open-weights reasoning model from NVIDIA, derived from Meta Llama-3.1-8B-Instruct through multi-phase post-training that combines supervised fine-tuning on math, code, reasoning and tool calling with reinforcement learning (RLOO and Online RPO). A detailed thinking on/off system prompt switches between fast non-reasoning chat and a reasoning mode; official benchmarks report MATH500 rising from 36.6% to 95.4% pass@1, AIME25 from 0% to 47.1%, and MBPP from 66.1% to 84.6% with reasoning on. With up to 128K context and a footprint that fits a single RTX GPU, it targets local and single-GPU deployment for AI agents, chatbots, RAG and instruction-following applications in English and code.

Llama-3.3 Nemotron Super 49B v1

Llama-3.3 Nemotron Super 49B v1 is an open reasoning model from NVIDIA, derived from Meta's Llama-3.3-70B-Instruct through Neural Architecture Search (NAS) and knowledge distillation. The 49B parameter design removes or replaces attention blocks and varies FFN ratios per block, cutting memory footprint enough to fit on a single H100 80GB GPU, with a context length of 128K tokens. Reasoning ON or OFF is switched via the system prompt, and multi-phase post-training (SFT plus RLOO and Online RPO reinforcement learning) strengthens math, code, RAG and tool calling. On official benchmarks it scores 96.6 on MATH500 and 58.4 on AIME25 with reasoning ON, 88.3 on Arena-Hard and 89.21 on IFEval with reasoning OFF. Released under the NVIDIA Open Model License for commercial use, it targets AI agents, chatbots and RAG applications.

Mistral Small 3.1 24B Base

Mistral Small 3.1 24B Base is the pretrained base checkpoint of Mistral Small 3.1, released by Mistral AI under the Apache 2.0 license. It is a 24B-parameter multimodal model with vision understanding, a 128k token context window, a Tekken tokenizer (131k vocabulary), and support for dozens of languages. Unlike the Instruct version, it ships without instruction post-training and is not usable as a chat model out of the box; Mistral positions it for community fine-tuning, domain experts in areas like legal, medical, or technical support, and as a foundation for reasoning models. On the official pretrain evals it scores 81.01% on MMLU, 56.03% on MMLU Pro, 37.50% on GPQA Main, and 59.27% on MMMU, ahead of Gemma 3 27B PT on most listed benchmarks. It runs on a single RTX 4090 or a Mac with 32GB of RAM, with vLLM as the recommended inference engine.

Mistral Small 3.1 24B Instruct

Mistral Small 3.1 24B Instruct is an open source multimodal LLM from Mistral AI, released under Apache 2.0 with 24B parameters. Building on Mistral Small 3, it adds vision understanding and a 128k token context window while keeping text performance, and delivers inference at about 150 tokens per second. It fits on a single RTX 4090 or a 32GB RAM Mac once quantized, and official benchmarks show it matching or beating Gemma 3 27B and GPT-4o mini across text, vision, multilingual and long context evaluations such as MMLU, DocVQA and RULER. Typical uses include fast conversational assistants, low-latency function calling in agentic workflows, image understanding like document verification and visual inspection, domain fine-tuning, and local inference on sensitive data. It is downloadable on Hugging Face, served via La Plateforme and Google Cloud Vertex AI, with NVIDIA NIM and Azure AI Foundry support to follow.

Gemini Robotics

Gemini Robotics is a vision-language-action (VLA) model from Google DeepMind, built on Gemini 2.0 and adding physical actions as a new output modality so it can control robots directly. It handles dexterous multi-step tasks such as origami folding and packing a snack into a Ziploc bag, follows everyday conversational instructions in different languages, and adapts to new objects, diverse instructions, and new environments across embodiments including the ALOHA 2 bi-arm platform, Franka arms, and the Apollo humanoid. Google DeepMind reports that it more than doubles performance on a comprehensive generalization benchmark compared with other state-of-the-art VLA models, and the team is partnering with Apptronik on next generation humanoid robots.

Gemini Robotics-ER

Gemini Robotics-ER (Embodied Reasoning) is a vision-language model from Google DeepMind, built on Gemini 2.0 for robotics developers. It adds advanced spatial understanding to Gemini, improving pointing and 3D detection by a large margin, and combines spatial reasoning with coding so robots can handle perception, state estimation, planning and code generation end to end, including two-finger grasp inference from a single image. Google DeepMind reports a 2x to 3x end-to-end success rate compared with Gemini 2.0. The model is available to trusted testers including Agile Robots, Agility Robots, Boston Dynamics and Enchanted Tools, and it pairs with the vision-language-action model Gemini Robotics, which directly controls robot actions.

Gemma 3

Gemma 3 is a family of lightweight, open models from Google DeepMind, built from the same research as Gemini 2.0 and offered in four sizes: 1B, 4B, 12B, and 27B. The 4B, 12B, and 27B variants are multimodal and take text and image input with a 128K-token context window (32K for the 1B model), and the family is pretrained on over 140 languages. Google reports that the 27B model ranks among the top open models on the LMArena human preference leaderboard while running on a single NVIDIA H100, and official quantized versions target phones, laptops, and workstations with support from NVIDIA, AMD ROCm, and gemma.cpp. Weights are available via Google AI Studio, Hugging Face, Kaggle, and Ollama, support function calling and structured output, and ship alongside ShieldGemma 2, a 4B image safety classifier.

hunyuan-video-super-resolution

hunyuan-video-super-resolution is the super resolution variant of Tencent Hunyuan video generation, an enhancement stage that upscales generated videos to 1080p resolution. Tencent describes its few-step super-resolution network as sharpening details, correcting distortions, and refining overall visual texture, and the module ships enabled by default in the official HunyuanVideo inference pipeline. It suits workflows that deliver generated clips at higher resolution, such as short video production, advertising assets, and post-production finishing.

hunyuan-video-voiceover

hunyuan-video-voiceover is an aggregator identifier for the voiceover and dubbing variant of Tencent Hunyuan's video generation family. The capability maps to the officially released HunyuanVideo line: with the image-to-video upgrade open sourced on Mar 6, 2025, the model fuses image, text, audio and pose inputs to control generated clips. Its audio-driven mode can read input text aloud through a choice of voice styles, or take an uploaded voice track and align mouth shapes precisely so the character speaks or sings with tone-fitting facial expressions. Typical use cases include virtual anchors, short-video narration, talking avatars and dubbed content.

2a

Ideogram 2a is a text-to-image model released in February 2025 as the fastest and most affordable option in the Ideogram lineup at launch. Optimized for graphic design and photography, it keeps Ideogram's signature text rendering and produces premium designs with text and photorealistic images in about 10 seconds, or 5 seconds with 2a Turbo, while images cost 50% less on both the website and API. It ships in the API as the V_2A and V_2A_TURBO versions and suits high-volume scenarios such as posters, marketing creatives, and ad layouts.

GPT-4.5

GPT-4.5 is OpenAI's largest and most capable chat model at release, launched on 2025-02-27 as a research preview (previously known by the codename Orion). It scales unsupervised learning rather than chain-of-thought reasoning, bringing broader world knowledge, higher EQ, more natural conversation, and a lower hallucination rate on SimpleQA. Official benchmarks show 71.4% on GPQA (science), 85.1% on MMMLU (multilingual), 74.4% on MMMU, and 38.0% on SWE-Bench Verified. In ChatGPT it supports web search, file and image uploads, and canvas; in the API it offers function calling, Structured Outputs, streaming, and image inputs with a 128,000 token context window at $75/$150 per million input/output tokens. It fits writing help, creative work, agentic planning, and multi-step coding workflows; the gpt-4.5-preview API model has since been deprecated in favor of gpt-4.1 and o3.

Wan2.1-I2V-Turbo

Wan2.1-I2V-Turbo is the fast tier of Alibaba Tongyi's Wan2.1 image-to-video series. It takes a first-frame image plus a prompt and generates a silent MP4 video at a fixed 30fps, available in 480P or 720P and in 3, 4, or 5 second durations, making it suited to marketing short clips, previews, and iterative creative work where speed and cost matter more than the Plus tier's fidelity. Wan2.1 itself is an open video foundation model suite covering text-to-video, image-to-video, and editing, and it was the first video model family to generate both Chinese and English text within video.

Alexa+

Alexa+ is Amazon's next-generation voice assistant announced on February 26, 2025. According to Anthropic's announcement, Claude models accessed through Amazon Bedrock help power Alexa+, with the two teams working closely for over a year, and Alexa+ also draws on Anthropic's safety capabilities such as advanced jailbreaking resistance. Alexa+ offers natural conversation, personalization, and agentic task execution, including reservations, grocery ordering, ticket reminders, and schedule management, with wider availability to follow.

Claude 3.7 Sonnet

Claude 3.7 Sonnet is Anthropic's large language model released on February 24, 2025, described by the company as its most intelligent model to date and the first hybrid reasoning model on the market. A single model handles both instant responses and extended thinking, where it reasons step by step with a visible chain of thought, and API developers can tune reasoning depth with a thinking budget within a 128K output limit. On official benchmarks it scores 70.3% on SWE-bench Verified with a high-compute scaffold and 63.7% without one, reports state-of-the-art results on TAU-bench for agentic tool use, and cuts unnecessary refusals by 45% compared to its predecessor. Pricing matches earlier Sonnet models at $3 per million input tokens and $15 per million output tokens, thinking tokens included. It is available across all Claude plans, the Claude Developer Platform, Amazon Bedrock, and Google Cloud's Vertex AI, and shipped alongside Claude Code, a command-line tool for agentic coding. Anthropic positions it especially for coding and front-end web development, as well as business tasks that benefit from step-by-step reasoning.

hunyuan-video-aries-i2v

HunyuanVideo-I2V is Tencent Hunyuan's open-source image-to-video generation framework built on top of the HunyuanVideo model. It takes a reference image plus a text prompt and generates videos up to 720P resolution and 129 frames (about 5 seconds). A token replace technique turns the input image into semantic tokens via a pretrained MLLM text encoder, concatenates them with video latent tokens for full-attention computation, and improves semantic alignment between the image, the caption, and the generated motion. Users can pick a stability mode for steadier output or a high-dynamic mode through flow-shift settings, and the first frame stays visually consistent with the input image. The repo also ships LoRA training code for custom effects and xDiT-based multi-GPU parallel inference that cuts 720P generation latency to about 337.58 seconds on 8 GPUs, a 5.64x speedup. It suits animating still images for marketing clips, social content, and VFX prototyping, with roughly 60GB of GPU memory required for 720P inference.

hunyuan-video-portrait

HunyuanPortrait is a diffusion-based portrait animation framework open sourced by Tencent Hunyuan. It takes a single portrait photo as the identity reference and a driving video as the motion reference, encoding facial expressions and head pose from the video into implicit control signals that are injected into a stabilized Stable Video Diffusion backbone through attention-based adapters. By decoupling identity from motion, it generates lifelike, temporally consistent and style-flexible animations, and the official cases cover portrait singing, acting and exaggerated expressions. The team states the method outperforms existing approaches in controllability and coherence; code and weights are released for academic use and run on a single 24 GB GPU.

FLUX 1.1 [pro] Ultra

FLUX 1.1 pro Ultra is the high-resolution tier of the FLUX 1.1 pro image model from Black Forest Labs, offered through the official BFL API. It generates images at up to 4 megapixels (4MP), four times the resolution of standard FLUX 1.1 pro, while keeping generation time at only about 10 seconds per image. Based on official performance benchmarks it stays over 2.5x faster than comparable high-resolution offerings without sacrificing prompt adherence. It fits workloads that need large, detailed output such as advertising visuals, print layouts, and photorealistic content.

FLUX.1 Canny [pro]

FLUX.1 Canny Pro is the pro variant of the Canny model in the FLUX.1 Tools suite by Black Forest Labs, built for structure-guided image generation. It extracts a Canny edge map from an input image and uses it as structure guidance alongside a text prompt, regenerating image content while the original outlines and composition stay intact. Black Forest Labs describes this approach as particularly effective for retexturing images, making it a fit for style changes, material and texture swaps, and controlled edits where the layout must hold. In the company's own benchmark evaluation, FLUX.1 Canny Pro ranked best in class among Canny-based structural guidance models, ahead of the open-weight dev variant and its LoRA versions available on Hugging Face.

FLUX.1 Depth [pro]

The pro version of FLUX.1 Depth, part of the FLUX.1 Tools suite by Black Forest Labs. It extracts a depth map from an input image and uses it as structure guidance to regenerate the image from a text prompt, keeping the core composition intact so edits remain text-guided. This form of structural conditioning is particularly effective for retexturing, such as changing materials or visual styles while preserving the spatial layout. According to the official announcement, FLUX.1 Depth outperforms proprietary models like Midjourney ReTexture, and the pro version offers higher output diversity than the dev variant. The model was served through the BFL API and partner platforms such as fal.ai and Replicate, and has since been deprecated in the BFL API with open weights kept for reference.