All 22 verified AI news entries from 2024, including 16 model releases, in date order with a link to each official source.
Showing 22 of 22 entries
DeepSeekModelsMajor
DeepSeek-V3 was a mixture-of-experts model with 37B of 671B parameters active, trained on 14.8T tokens in 2.788M H800 GPU hours, with open weights that allowed commercial use.
OpenAIModelsMajor
On the last day of its 12 Days event, OpenAI previewed o3 and o3-mini and opened early access for safety researchers. ARC Prize verified 75.7% on ARC-AGI at its cost limit and 87.5% with far more compute.
Image: OpenAI
Google DeepMindCreative, Important
Google said human raters preferred Veo 2 over leading video models. It could output up to 4K, rolled out through VideoFX with a waitlist, and carried SynthID watermarks. An updated Imagen 3 shipped the same day.
Image: Google DeepMind
Google DeepMindModelsMajor
Gemini 2.0 Flash outperformed Gemini 1.5 Pro on key benchmarks at about twice the speed, added native tool use and image and audio output, and launched with the Multimodal Live API.
Image: Google DeepMind
OpenAICreative, Important
Sora became a standalone product at sora.com, generating up to 1080p videos of up to 20 seconds from text, images or video, included in ChatGPT Plus with higher limits on Pro.
Image: OpenAI
AnthropicModelsMajor
The upgraded Claude 3.5 Sonnet rose from 33.4% to 49.0% on SWE-bench Verified at the same price, and became the first frontier model offered in public beta to operate a computer through screenshots, mouse and keyboard.
Image: Anthropic
OpenAIVoice, Important
The Realtime API let paid developers build low-latency voice conversations on GPT-4o over a single WebSocket connection, with function calling, at about $0.06 per minute of audio input and $0.24 per minute of output.
Image: OpenAI
Alibaba Cloud (Qwen)Models, Important
Qwen2.5 was pretrained on about 18 trillion tokens, supported 128K context and more than 29 languages, and came with coding and math variants, mostly under Apache 2.0.
OpenAIModelsMajor
The o1 series spent more time reasoning before answering. OpenAI said its next update scored 83% on an IMO qualifying exam versus 13% for GPT-4o. o1-preview and o1-mini launched in ChatGPT and the API.
Image: OpenAI
Black Forest LabsCreative, Important
FLUX.1 came in three variants: pro via API, dev with open weights for non-commercial use, and schnell under Apache 2.0. The company also announced a $31 million seed round led by Andreessen Horowitz.
Image: Black Forest Labs
MetaModelsMajor
Llama 3.1 came in 8B, 70B and 405B sizes with a 128K context and eight languages. Meta said the 405B was competitive with GPT-4o and Claude 3.5 Sonnet and allowed its outputs to train other models.
OpenAIModels, Important
GPT-4o mini scored 82% on MMLU, beat GPT-3.5 Turbo and other small models on OpenAI's benchmarks, and cost $0.15/$0.60 per 1M tokens with a 128K context window.
Image: OpenAI
GoogleModels, Important
Google said Gemma 2 27B competed with models more than twice its size and ran at full precision on a single H100 GPU, under a license that allowed commercial use.
Image: Google
AnthropicModelsMajor
Claude 3.5 Sonnet ran twice as fast as Claude 3 Opus at $3/$15 per 1M tokens, solved 64% of problems in Anthropic's agentic coding eval, and launched with Artifacts on claude.ai.
Image: Anthropic
RunwayCreative, Important
Gen-3 Alpha was trained jointly on video and images, added finer control over timing and more realistic people, and was set to power Runway's text-to-video and image-to-video tools.
Image: Runway
OpenAIModelsMajor
GPT-4o matched GPT-4 Turbo on English text and code, responded to audio in about 320 milliseconds on average, came to ChatGPT's free tier, and cost half as much as GPT-4 Turbo in the API.
Image: OpenAI
MetaModels, Important
Llama 3 used about seven times more training data than Llama 2, a new 128K-token vocabulary and wide cloud availability, and it powered the Meta AI assistant.
SunoCreative, Important
Suno called v3 its first model capable of radio-quality music. It made songs of up to two minutes in seconds, was open to free users, and carried an inaudible watermark.
Image: Suno
xAIModels, Important
xAI published the raw Grok-1 base checkpoint, a 314B mixture-of-experts model with about 25% of weights active per token, under Apache 2.0.
AnthropicModelsMajor
Claude 3 brought vision to all three models, a 200K context window and three price tiers, from Haiku at $0.25 per 1M input tokens to Opus at $15.
Google DeepMindModelsMajor
Gemini 1.5 Pro used a mixture-of-experts design, performed close to Gemini 1.0 Ultra, and offered a 1M-token context in private preview, enough for about an hour of video or 700,000 words.
Image: Google DeepMind
OpenAICreativeMajor
Sora generated videos of up to one minute from text. OpenAI shared it with red teamers and a group of artists and filmmakers rather than releasing it publicly.
Image: OpenAI
