Releases

Microsoft releases Phi-4 Mini, a compact long-context language model with multilingual and function-calling capabilities.

Anthropic released Claude 3.7 Sonnet, its first hybrid reasoning model with standard and extended-thinking modes.

Hugging Face released SmolVLM2 at 256M, 500M, and 2.2B parameters with support for video understanding as well as images and text.

xAI unveiled Grok 3 and Grok 3 mini with reasoning modes, DeepSearch and Think capabilities trained on Colossus.

Perplexity released R1 1776, an open-weight post-trained version of DeepSeek-R1 intended to reduce censorship while retaining reasoning capability.

Mistral AI released Mistral Saba, a 24B regional language model optimized for Arabic and languages of the Middle East and South Asia.

Perplexity released a faster in-house Sonar model built on Llama 3.3 70B and served through Cerebras for Pro search.

Google releases Gemini 2.0 Flash-Lite in public preview as its lowest-cost Gemini 2.0 model.

Google releases an experimental Gemini 2.0 Pro for coding and complex prompts with a two-million-token context window.

Gemini 2.0 Flash becomes generally available through the Gemini API, Google AI Studio and Vertex AI.

Moonshot released Moonlight, a 16B MoE model used to evaluate Muon optimization at scale.

OpenAI released o3-mini, a cost-efficient reasoning model optimized for science, mathematics, and coding.

Mistral AI released Mistral Small 3, a 24B latency-optimized open-weight model.

Perplexity introduced Sonar Reasoning and Sonar Reasoning Pro for search-grounded analytical queries.

DeepSeek released scaled 1B and 7B Janus-Pro checkpoints with improved understanding and image generation.

Hugging Face released a 500M vision-language model for low-memory multimodal inference.

Hugging Face released a 256M vision-language model small enough for highly constrained devices while supporting image-and-text tasks.

Moonshot released Kimi k1.5, a multimodal reasoning model trained with reinforcement learning.

Moonshot introduced its open k-series reasoning work through the Kimi k1.5 release.

DeepSeek released the Llama 70B-based distilled R1 reasoning checkpoint.

DeepSeek released the Llama 8B-based distilled R1 reasoning checkpoint.

DeepSeek released the Qwen 14B-based distilled R1 reasoning checkpoint.

DeepSeek released the Qwen 1.5B-based distilled R1 reasoning checkpoint.

DeepSeek released the Qwen 7B-based distilled R1 reasoning checkpoint.

DeepSeek released the Qwen 32B-based distilled R1 reasoning checkpoint.

DeepSeek published R1 under the MIT License and made it available on web and API.

DeepSeek published its pure reinforcement-learning reasoning checkpoint.

Mistral AI released an updated Codestral model with improved code generation and a 256K context window.

NVIDIA launched Cosmos with open world foundation models, tokenizers, guardrails and data-processing tools for physical AI.

DeepSeek released DeepSeek-V3 weights and upgraded the deepseek-chat API.

Answer.AI and Hugging Face released ModernBERT, a modernized encoder-only model family for retrieval and classification workloads.

IBM released open embedding models for enterprise retrieval and semantic search.

IBM expanded Granite with improved language, embedding and safety models.

Google DeepMind introduces Veo 2 with improved realism, physics and camera controls for video generation.

DeepSeek published its second-generation mixture-of-experts vision-language models.

Meta released a behavioral foundation model for controlling virtual humanoid agents.

Microsoft releases Phi-4, a 14-billion-parameter small language model specializing in complex reasoning.

Google releases experimental Gemini 2.0 Flash with native tool use, multimodal output and real-time streaming capabilities.

The final V2.5 update improved mathematics, coding, writing, file handling and web-page summarization.

xAI launched the autoregressive Aurora image-generation model on X, adding photorealistic generation and multimodal image input.

Meta released a multilingual 70B instruction model with performance comparable to larger Llama 3.1 models.

Google releases PaliGemma 2, expanding the open vision-language family with new model sizes and resolutions.

OpenAI released the full o1 reasoning model in ChatGPT and announced related API improvements.

AWS released its first Nova video-generation model.

AWS released its Nova image-generation and editing model.

AWS released a capable multimodal Nova model for enterprise and agentic workloads.

AWS released a low-cost multimodal Nova model through Amazon Bedrock.

AWS released the text-only low-latency Nova model through Amazon Bedrock.

Cohere introduced Rerank 3.5 with stronger reasoning and multilingual retrieval over complex enterprise data.

Amazon released a faster and more memory-efficient zero-shot forecasting model family.

Hugging Face released SmolVLM, a compact 2B vision-language model designed to run with very low memory.

Mistral AI released Pixtral Large, a 124B multimodal model for image and document understanding.

DeepSeek released JanusFlow using rectified flow for image generation.

Scale released Defense Llama, a Llama 3-based model fine-tuned for controlled U.S. national-security use cases inside Scale Donovan.

Anthropic released Claude 3.5 Haiku as a faster model with substantially improved intelligence and coding performance.

Hugging Face released the second generation of its 135M, 360M, and 1.7B compact language models with stronger instruction following.

Cohere Labs released Aya Expanse 32B as the larger Aya Expanse multilingual model.

Cohere Labs released Aya Expanse 8B with improved multilingual generation across 23 languages.

Anthropic released an upgraded Claude 3.5 Sonnet with major coding and tool-use improvements.

IBM released its Granite 3.0 8B model with enterprise and agentic capabilities under Apache 2.0.

IBM released its compact Granite 3.0 2B enterprise language model under Apache 2.0.

DeepSeek released Janus 1.3B for multimodal understanding and image generation.

Mistral AI launched the Ministral 3B and 8B models for edge, on-device and local use.

Meta published improved SAM 2.1 checkpoints and developer tooling.

Meta released a 90B multimodal model for image understanding.

Meta released an 11B multimodal model for image understanding.

Meta released a compact 1B text model for on-device use.

Meta released a compact 3B text model for on-device use.

Mistral AI released Pixtral 12B, its first open-weight multimodal model.

OpenAI released its first o-series reasoning models, o1-preview and o1-mini.

DeepSeek merged its general and coding capabilities into V2.5 and published weights.

The August Command R+ release delivered roughly 50% higher throughput and 25% lower latency with the same hardware footprint.

Cohere updated Command R with improved tool decisions, instruction following, structured-data analysis and robustness.

Hugging Face released IDEFICS3, a more efficient multimodal architecture that became the basis for the SmolVLM family.

Microsoft releases Phi-3.5 Vision for multimodal image and text understanding.

Microsoft releases Phi-3.5 MoE, a mixture-of-experts small language model.

Microsoft releases Phi-3.5 Mini with multilingual support and a long context window.

DeepSeek released an improved formal theorem prover using reinforcement learning and search.

Google begins broad availability of Imagen 3 through ImageFX and later expands access through Gemini and Vertex AI.

xAI released Grok-2 and Grok-2 mini in beta on X and announced enterprise API availability.

Amazon updated its image model with image conditioning, subject consistency and background-removal capabilities.

Google releases ShieldGemma, open safety classifiers built on Gemma 2 for filtering model inputs and outputs.

Meta released SAM 2 for promptable object segmentation in both images and video.

DeepSeek upgraded the deepseek-coder API to the July Coder V2 checkpoint.

Mistral AI released Mistral Large 2, a 123B flagship model with a 128K context window and improved multilingual and coding performance.

Meta released a classifier for detecting prompt injection and jailbreak attempts.

Meta released a multilingual Llama 3.1-based safeguard model.

Meta released its 405B-parameter open-weights frontier model.

Meta released Llama 3.1 70B with a 128K context window and multilingual support.

Meta released Llama 3.1 8B with a 128K context window and multilingual support.

OpenAI released GPT-4o mini as a lower-cost small multimodal model.

Hugging Face released compact fully open language models at 135M, 360M, and 1.7B parameter scales.

Mistral AI released Codestral Mamba, an open code model using the Mamba 2 architecture.

Mistral AI released Mathstral 7B, an open model specialized for mathematical reasoning.

Baidu released a faster and more capable version of ERNIE 4.0.

DeepSeek updated its chat model with stronger code, math, reasoning and role-play performance.

Google releases Gemma 2, led by a 27B model with improved performance and efficiency.

EvolutionaryScale, founded by former Meta researchers, released ESM-3, a generative protein model originating from Meta's ESM program.

Anthropic released Claude 3.5 Sonnet with stronger reasoning, coding, writing, and vision performance.

DeepSeek released 16B and 236B mixture-of-experts coding models with 128K context.