All tags
Topic: "model-compression"
DeepSeek-OCR finds vision models can decode 10x more efficiently with ~97% accuracy of text-only, 33/200k pages/day/A100
deepseek-ocr deepseek3b-moe-a570m veo-3.1 deepseek-ai google-deepmind krea ocr vision multimodality model-compression long-context model-architecture video-generation autoregressive-models model-efficiency precision-editing karpathy teortaxestex reach_vb _akhaliq eliebakouch vikhyatk demishassabis
As ICCV 2025 begins, DeepSeek releases a novel DeepSeek-OCR 3B MoE vision-language model that compresses long text as visual context with high accuracy and efficiency, challenging traditional tokenization approaches. The model achieves ~97% decoding precision at <10× compression and processes up to ~33M pages/day on 20 A100-40G nodes, outperforming benchmarks like GOT-OCR2.0. Discussions highlight the potential for unlimited context windows and tokenization-free inputs, with contributions from @karpathy, @teortaxesTex, and others. In video generation, google-deepmind's Veo 3.1 leads community benchmarks with advanced precision editing and scene blending, while Krea open-sources a 14B autoregressive video model enabling realtime long-form generation at ~11 FPS on a single B200 GPU.
not much happened today
embeddinggemma qwen-2.5-coder minicpm-v-4.5 gpt-4o gemini-2.0-pro google-deepmind hugging-face jina-ai lighton microsoft stanford openai ollama weaviate langchain llamaindex embeddings retrieval-augmented-generation quantization multilingual-models on-device-ai semantic-search contrastive-learning dataset-release vision multimodality video-generation text-to-speech optimizer-benchmarking training-recipes model-compression video-token-compression fine-tuning osanseviero _philschmid tomaarsen ollama weaviate_io lusxvr andimarafioti thibaudfrere _akhaliq clementdelangue gordonwetzstein konstmish wen_kaiyue percyliang
Google DeepMind released EmbeddingGemma (308M), a small multilingual embedding model optimized for on-device retrieval-augmented generation and semantic search, supporting over 100 languages and running efficiently with quantization and EdgeTPU latency under 15ms. Jina AI introduced new code-focused embedding models (0.5B/1.5B) with GGUF quantization, achieving state-of-the-art retrieval across multiple languages and tasks. LightOn demonstrated large-scale retrieval training without distillation using contrastive training on billions of passages. Hugging Face released the FineVision dataset with 17.3M images and 9.5B answer tokens for vision-language model training, showing significant benchmark improvements. The MiniCPM-V 4.5 (8B) multimodal model reported surpassing GPT-4o and Gemini-2.0 Pro on OpenCompass benchmarks with innovative video token compression. Microsoft’s VibeVoice TTS and Stanford’s Mixture-of-Contexts video generation also featured. Additionally, a Stanford study benchmarked optimizers like Muon, Soap, Mars, and Sophia, finding diminishing speedups over AdamW at larger scales but advantages at smaller scales. The new ChatGPT branching feature was noted for its simplicity and popularity. "Everyone's a decacorn now."
Not much happened today
mistral-small-3.2 magenta-realtime afm-4.5b llama-3 openthinker3-7b deepseek-r1-distill-qwen-7b storm qwen2-vl gpt-4o dino-v2 sakana-ai mistral-ai google arcee-ai deepseek-ai openai amazon gdm reinforcement-learning chain-of-thought fine-tuning function-calling quantization music-generation foundation-models reasoning text-video model-compression image-classification evaluation-metrics sama
Sakana AI released Reinforcement-Learned Teachers (RLTs), a novel technique using smaller 7B parameter models trained via reinforcement learning to teach reasoning through step-by-step explanations, accelerating Chain-of-Thought learning. Mistral AI updated Mistral Small 3.2 improving instruction following and function calling with experimental FP8 quantization. Google Magenta RealTime, an 800M parameter open-weights model for real-time music generation, was released. Arcee AI launched AFM-4.5B, a sub-10B parameter foundation model extended from Llama 3. OpenThinker3-7B was introduced as a new state-of-the-art 7B reasoning model with a 33% improvement over DeepSeek-R1-Distill-Qwen-7B. The STORM text-video model compresses video input by 8x using Mamba layers and outperforms GPT-4o on MVBench with 70.6%. Discussions on reinforcement learning algorithms PPO vs. GRPO and insights on DINOv2's performance on ImageNet-1k were also highlighted. "A very quiet day" in AI news with valuable workshops from OpenAI, Amazon, and GDM.
Grok 3 & 3-mini now API Available
grok-3 grok-3-mini gemini-2.5-flash o3 o4-mini llama-4-maverick gemma-3-27b openai llamaindex google-deepmind epochairesearch goodfireai mechanize agent-development agent-communication cli-tools reinforcement-learning model-evaluation quantization-aware-training model-compression training-compute hybrid-reasoning model-benchmarking
Grok 3 API is now available, including a smaller version called Grok 3 mini, which offers competitive pricing and full reasoning traces. OpenAI released a practical guide for building AI agents, while LlamaIndex supports the Agent2Agent protocol for multi-agent communication. Codex CLI is gaining traction with new features and competition from Aider and Claude Code. GoogleDeepMind launched Gemini 2.5 Flash, a hybrid reasoning model topping the Chatbot Arena leaderboard. OpenAI's o3 and o4-mini models show emergent behaviors from large-scale reinforcement learning. EpochAIResearch updated its methodology, removing Maverick from high FLOP models as Llama 4 Maverick training compute drops. GoodfireAI announced a $50M Series A for its Ember neural programming platform. Mechanize was founded to build virtual work environments and automation benchmarks. GoogleDeepMind's Quantisation Aware Training for Gemma 3 models reduces model size significantly, with open source checkpoints available.
not much happened today
gpt-4.5 claude-3.7-sonnet deepseek-r1 smolagents-codeagent gpt-4o llama-3-8b tinyr1-32b-preview r1-searcher forgetting-transformer nanomoe openai deepseek hugging-face mixture-of-experts reinforcement-learning kv-cache-compression agentic-ai model-distillation attention-mechanisms model-compression minimax model-pretraining andrej-karpathy cwolferesearch aymericroucher teortaxestex jonathanross321 akhaliq
The AI news recap highlights several key developments: nanoMoE, a PyTorch implementation of a mid-sized Mixture-of-Experts (MoE) model inspired by Andrej Karpathy's nanoGPT, enables pretraining on commodity hardware within a week. An agentic leaderboard ranks LLMs powering smolagents CodeAgent, with GPT-4.5 leading, followed by Claude-3.7-Sonnet. Discussions around DeepSeek-R1 emphasize AI model commoditization, with DeepSeek dubbed the "OpenAI of China." Q-Filters offer a training-free method for KV cache compression in autoregressive models, achieving 32x compression with minimal perplexity loss. The PokéChamp minimax language agent, powered by GPT-4o and Llama-3-8b, demonstrates strong performance in Pokémon battles. Other notable models include TinyR1-32B-Preview with Branch-Merge Distillation, R1-Searcher incentivizing search capability via reinforcement learning, and the Forgetting Transformer using a Forget Gate in softmax attention. These advancements reflect ongoing innovation in model architectures, compression, reinforcement learning, and agentic AI.