New AI Models

Nova Pro

Nova Pro is Amazon's mid tier multimodal understanding model in the Nova family, offered through Amazon Bedrock. It accepts text, image, and video input with a 300K token context window, and AWS positions it as the best combination of accuracy, speed, and cost in the lineup. According to the AWS announcement, it achieves state-of-the-art results on TextVQA for visual question answering and VATEX for video understanding, and it is built for RAG, function calling, and agentic workflows that call APIs and tools, with support for over 200 languages. Typical use cases include document and video analysis, working with codebases, and building agents; it can also act as a teacher model for distilling custom Nova Micro or Lite variants.

v4

Suno v4 is an AI music generation model released in November 2024, first available in beta to Pro and Premier users. Compared with earlier versions it delivers cleaner audio, clearer vocals, and sharper lyric articulation, along with more dynamic song structures. It introduces Remaster, which upgrades tracks made with older models to v4 quality, adds the ReMi lyrics assist model, and brings v4 upgrades to Covers and Personas. It suits songwriters and creators polishing demos, refreshing earlier songs, and keeping a consistent style across tracks.

Pixtral Large

Pixtral Large is a 124B open-weights multimodal model from Mistral AI, built on top of Mistral Large 2 with a 123B multimodal decoder and a 1B parameter vision encoder. It understands documents, charts and natural images while keeping the text performance of Mistral Large 2, and supports a 128K context window that fits at least 30 high-resolution images. Mistral reports state-of-the-art results on MathVista (69.4%), DocVQA and VQAv2, with ChartQA and DocVQA scores surpassing GPT-4o and Gemini-1.5 Pro, and stronger results than Claude 3.5 Sonnet (new), Gemini-1.5 Pro and GPT-4o (latest) on MM-MT-Bench. Released under the Mistral Research License for research and educational use with a commercial license option, it is available on le Chat, via the API as pixtral-large-latest and as downloadable weights, suiting document parsing, chart analysis and multilingual OCR workloads.

Claude 3.5 Haiku

Claude 3.5 Haiku is Anthropic's fast, cost-efficient model in the Claude 3.5 family, announced on October 22, 2024 and available from November 4, 2024 via the Anthropic API, Amazon Bedrock, and Google Cloud Vertex AI. It runs at the speed tier of the previous Claude 3 Haiku while surpassing the larger Claude 3 Opus on many intelligence benchmarks, scoring 40.6% on SWE-bench Verified, with improved instruction following, more accurate tool use, and low latency. It offers a 200K token context window and is priced at $0.80 per million input tokens and $4 per million output tokens, fitting user-facing products, specialized sub-agent tasks, personalized experiences over large datasets, and latency-sensitive coding scenarios such as high-frequency code completion.

Ministral 8B Instruct

Ministral 8B Instruct is an 8B instruct-tuned language model from Mistral AI's les Ministraux family, released in October 2024 for on-device and edge use cases. Built as a dense Transformer with interleaved sliding-window attention, it supports a 128k context window, function calling, and a 131k vocabulary with the V3-Tekken tokenizer. Mistral's official benchmarks report it outperforming same-size instruct models such as Llama 3.1 8B Instruct and Gemma 2 9B Instruct, with 76.8 HumanEval pass@1, 54.5 Math maj@1, 70.9 Arena Hard and 8.3 MTBench. Typical uses include privacy-first local inference, offline smart assistants, on-device translation, and input parsing or task routing in multi-step agentic workflows. It is served via the ministral-8b-latest API and its weights are released under the Mistral Research License for research use.

FLUX 1.1 [pro]

FLUX 1.1 pro is Black Forest Labs' upgraded flagship text-to-image model, launched on the BFL API in October 2024. Per the official announcement, it improves image quality, prompt adherence, and output diversity while running six times faster than the original FLUX.1 pro baseline, or three times faster than the updated FLUX.1 pro. Tested in the Artificial Analysis text-to-image arena under the codename blueberry, it reached the highest overall ELO score on the leaderboard at the time. The model is natively set up for fast ultra high-resolution output up to 2k, with that high-resolution mode opening on the API later, and it is listed at 4 US cents per image, also available through partners such as Together.ai, Replicate, fal.ai, and Freepik. It fits production workloads that need fast, faithful, high-quality image generation via API.

FLUX.1 [dev]

FLUX.1 [dev] is Black Forest Labs' open-weight, guidance-distilled 12 billion parameter rectified flow transformer for text-to-image generation. Per the official model card, its output quality ranks second only to the API-only FLUX.1 [pro], with prompt following that matches closed source alternatives, and guidance distillation makes sampling more efficient. The weights are published on Hugging Face under the FLUX.1 [dev] Non-Commercial License: research, personal, and artistic use are welcome, while commercial deployment of the model requires a separate license from Black Forest Labs. It fits research, fine-tuning experiments, and local workflows built on ComfyUI or diffusers.

FLUX.1 [pro]

FLUX.1 pro is the flagship, closed source member of the FLUX.1 text-to-image family from Black Forest Labs. Unlike the open weights dev variant, it is distributed only through the official BFL API and partner platforms such as Replicate, fal.ai and Together.ai. Official positioning puts it at the top of the series for prompt adherence, visual quality, image detail and output diversity, with API level controls for resolution and content moderation. Around October 2024 it became generally available via the BFL API at 5 cents per image and was updated to generate the same output two times faster. It fits production workloads that need reliable prompt following and consistent commercial quality, such as marketing visuals, product imagery and design workflows. It was later succeeded by FLUX1.1 pro, which further improved image quality, prompt adherence and diversity.

Llama 3.1 Nemotron 70B Instruct

Llama 3.1 Nemotron 70B Instruct is a 70B large language model from NVIDIA, post-trained from Llama 3.1 70B Instruct to improve the helpfulness of its responses. NVIDIA aligned it with RLHF, specifically REINFORCE, using the Llama-3.1-Nemotron-70B-Reward model and HelpSteer2-Preference prompts. The official model card reports Arena Hard 85.0, AlpacaEval 2 LC 57.6, and GPT-4-Turbo MT-Bench 8.98, ranking first on all three automatic alignment benchmarks as of 1 Oct 2024, ahead of GPT-4o and Claude 3.5 Sonnet, plus a Chatbot Arena Elo of 1267. It accepts up to 128k input tokens, ships under the NVIDIA Open Model License, and fits general instruction following and chat use cases; NVIDIA notes it is not tuned for specialized domains such as math.

Llama 3.2 11B Instruct

Llama 3.2 11B Instruct (Llama 3.2-Vision 11B) is Meta's open instruction-tuned multimodal model released on September 25, 2024. Built on the Llama 3.1 text-only model, it adds a separately trained vision adapter whose cross-attention layers feed image encoder representations into the LLM, taking text and images as input and generating text, with a 128K token context. It is optimized for visual recognition, image reasoning, captioning, and assistant-like chat about images, covering use cases such as VQA, DocVQA, chart and diagram understanding, image captioning, and visual grounding. Official benchmarks for the 11B model report DocVQA 88.4 (ANLS), ChartQA 83.4, AI2 Diagram 91.1, VQAv2 75.2, and MMMU 50.7 (CoT), and Meta states the Vision models outperform many available open source and closed multimodal models on common industry benchmarks. Image plus text input supports English only, while text-only tasks cover eight official languages. It fits document parsing, image-aware assistants, and synthetic data generation under the Llama 3.2 Community License.

Llama 3.2 3B Instruct

Llama 3.2 3B Instruct is Meta's instruction-tuned text model with 3.21B parameters, released on September 25, 2024 and positioned for on-device and edge use cases such as phones and laptops. It supports a 128K token context, officially covers 8 languages (English, German, French, Italian, Portuguese, Hindi, Spanish, Thai), and was aligned with SFT, rejection sampling and DPO plus distillation from Llama 3.1 models, targeting assistant-like chat, knowledge retrieval and summarization. Meta reports MMLU 63.4, IFEval 77.4, ARC-Challenge 78.6 and GSM8K (CoT) 77.7 for the bf16 model; a 4-bit SpinQuant build decodes at 19.7 tokens/sec on a OnePlus 12 via ExecuTorch.

Llama 3.2 90B Instruct

Llama 3.2 90B Instruct is Meta's open-weight, instruction-tuned multimodal model released on September 25, 2024. It builds on the Llama 3.1 text-only model with a separately trained vision adapter that uses cross-attention layers, accepting multilingual text plus images as input and producing text output over a 128K token context. Meta optimized it for visual recognition, image reasoning, captioning, and answering general questions about an image, with use cases spanning VQA, DocVQA, image captioning, image-text retrieval, and visual grounding. Official benchmarks for the instruction-tuned 90B variant include 60.3 on MMMU (CoT), 85.5 on ChartQA, 90.1 on DocVQA, and 86.0 on MMLU (CoT). Image plus text applications support English only, while text-only tasks cover eight languages including English, German, French, Spanish, and Hindi.

Mistral Small

Mistral Small v24.09 is Mistral AI's enterprise-grade small model from the September 2024 release, the upgrade to Mistral Small v24.02 served by the mistral-small-latest API endpoint. With 22B parameters, it sits between Mistral NeMo 12B and Mistral Large 2 as a cost-efficient option for translation, summarization, sentiment analysis, and other workloads that do not require a full-blown general purpose model. Mistral reports significant gains over the previous version in human alignment, reasoning, and code, with benchmark charts covering GSM8K, MATH, HumanEval, HumanEval Plus, MT Bench, and WildBench. Priced at $0.2 per M input tokens and $0.6 per M output tokens, an 80% cut from the prior rates, it runs on la Plateforme and cloud partners such as Azure AI Studio, Amazon Bedrock, and Google Vertex AI, and is released under the Mistral Research License for non-commercial self-deployment with tools like vLLM.

Pixtral-12B

Pixtral-12B is Mistral AI's first open multimodal model, released under Apache 2.0. Per the official announcement, it pairs a 12B parameter multimodal decoder based on Mistral Nemo with a newly trained 400M parameter vision encoder, ingests images at their native resolution and aspect ratio, and handles any number of images within a 128K token context window. It is built for documents, charts, diagrams and natural images, scoring 52.5% on the MMMU reasoning benchmark, and serves as a drop-in replacement for Mistral Nemo 12B while keeping strong text-only instruction following, coding and math performance. Good fits include document question answering, chart understanding, multimodal reasoning and image-to-code workflows via La Plateforme, Le Chat or Hugging Face weights.

o1

o1 is OpenAI's formal reasoning model family, released on 2024-09-12 as o1-preview for ChatGPT and trusted API users. Trained with large-scale reinforcement learning, it generates a long internal chain-of-thought before answering, learning to spot and fix its own mistakes, break hard steps into simpler ones, and switch approaches when one stalls. Official benchmarks show 74% pass@1 (83% cons@64) on AIME 2024, an 89th percentile Codeforces Elo of 1673, accuracy above PhD-level experts on GPQA-diamond, and 78.2% on MMMU. It fits complex math, science, and coding work such as competition problems and research-level questions.

o1-mini

o1-mini is a smaller, cost-efficient reasoning model from OpenAI, released on September 12, 2024 alongside o1-preview. It was trained with the same high-compute reinforcement learning pipeline as the full o1 model, with pretraining focused on STEM. OpenAI reports that it scores 70.0% on AIME math competition problems, close to o1 at 74.4% and far above o1-preview at 44.6%, and reaches 1650 Elo on Codeforces, near o1's 1673, at 80% lower API pricing than o1-preview. It also outperforms GPT-4o on GPQA and MATH-500 and generates reasoning roughly 3 to 5 times faster than o1-preview. Its world knowledge is narrower, with non-STEM factual accuracy comparable to GPT-4o mini, so it suits applications that need fast, affordable math and coding reasoning rather than broad general knowledge.

Hailuo 1.0

Hailuo 1.0 is MiniMax's first AI-native video generation model, launched on Hailuo AI on August 31, 2024 under the internal code name video-01. It supports both text-to-video and image-to-video generation, producing 720p clips at 25fps with a length of up to 6 seconds. According to the official release, the model delivers high compression rates, strong responsiveness to prompts, and diverse styles, with cinematic camera movement effects. It fits short-form video creation, storyboard previews, and creative exploration from text or a reference image.

Hailuo 1.0-Director

Hailuo 1.0-Director is the Director variant of the Hailuo video 1.0 family from MiniMax. Creators write natural language prompts with camera movement instructions such as pan, push in, zoom and follow, and the model executes the shots accordingly. MiniMax's official announcement says the Director models reduce movement randomness and are fine-tuned for stronger prompt adherence, so each frame follows the creative intent. It fits text-to-video and image-to-video workflows for short films, ads and storyboards that need precise camera control, and it is available on hailuoai.video and the MiniMax API platform.

Hailuo 1.0-Live

Hailuo 1.0-Live (video-01-Live) is the Live variant of MiniMax's Hailuo video family, whose first AI-native video generation model video-01 was officially released on August 31, 2024. The base model generates 720p 25fps clips up to 6 seconds and offers both text-to-video and image-to-video modes, with cinematic camera movement effects, high compression efficiency, strong responsiveness to text prompts, and native high resolution and high frame rate comparable to cinematic quality. MiniMax's official materials describe the Live variant as breathing life into static illustrations and turning them into dynamic works, which makes it a fit for anime, manga, and illustration creators who need expressive character motion from a single image. It is offered through MiniMax's video ecosystem and partner platforms such as Freepik.

Phi-3.5-mini-instruct

Phi-3.5-mini-instruct is Microsoft's 3.8B parameter open instruct model (MIT license), released in August 2024 as an update to Phi-3-mini. This dense decoder-only Transformer supports a 128K token context and was further trained on multilingual data covering 23 languages, with Microsoft citing 25% to 50% gains in Arabic, Dutch, Finnish, Polish, Thai and Ukrainian over its predecessor. The official model card reports 69 on MMLU (5-shot), 86.2 on GSM8K, 62.8 on HumanEval and 55.4 on Multilingual MMLU. Microsoft targets it at memory and compute constrained environments, latency bound scenarios, and reasoning tasks in code, math and logic, such as long document summarization, document QA and information retrieval, with RAG recommended for knowledge intensive multilingual workloads.