DeepSeek's New V3 and R1 Models: Efficiency Breakthroughs and Cost Savings

Model Releases and Updates

DeepSeek V3 and R1 Models: DeepSeek has released detailed pricing and system architecture information for their V3 and R1 models, showing significantly lower costs compared to competitors. Their inference system demonstrates remarkable efficiency with a theoretical cost-profit ratio of 545%.

• New Deepseek report: $1.10 per 1M output tokens for v3, $2.19 per 1M tokens for r1
  • https://x.com/Yuchenj_UW/status/1895228204564259067/photo/1
• Day 6: One More Thing, DeepSeek-V3/R1 Inference System Overview
  • https://github.com/deepseek-ai/open-infra-index/blob/main/202502OpenSourceWeek/day6onemorethingdeepseekV3R1inferencesystemoverview.md
• China's DeepSeek claims theoretical cost-profit ratio of 545% per day
  • https://finance.yahoo.com/news/chinas-deepseek-claims-theoretical-cost-121658741.html

TinyR1-32B: Efficient Model with SuperDistillation: Qihoo 360 has released TinyR1-32B-Preview, achieving near-R1 performance with only 5% of the parameters through a process called SuperDistillation.

• TinyR1-32B-Preview: SuperDistillation Achieves Near-R1 Performance with Just 5% of Parameters
  • https://huggingface.co/qihoo360/TinyR1-32B-Preview

Phi-4-mini Bug Fixes and GGUF Conversion: Llama.cpp has added support for Microsoft's Phi-4 mini model and fixed several tokenization issues, making the model more accessible for local deployment.

• Phi-4-mini Bug Fixes + GGUFs
  • https://huggingface.co/unsloth/Phi-4-mini-instruct-GGUF
  • https://huggingface.co/collections/unsloth/phi-4-all-versions-677eecf93784e61afe762afa
  • https://huggingface.co/unsloth/Phi-4-mini-instruct-unsloth-bnb-4bit

Open-Source DeepResearch Implementation: A new open-source project called Search-R1 has been released, reproducing the methods of DeepSeek-R1 for training reasoning and searching interleaved LLMs.

• The first real open source DeepResearch attempt I've seen
  • https://github.com/volcengine/verl
  • https://github.com/PeterGriffinJin/Search-R1/tree/main

Voice AI Breakthroughs

Sesame Real-Time Voice Chat Model: Sesame AI Labs has released a real-time, low-latency voice chat model available in three sizes, designed to create natural conversation with human-like qualities.

• A new realtime voice chat: Sesame
  • https://www.sesame.com/research/crossingtheuncannyvalleyof_voice#demo
• Finally, a real-time low-latency voice chat model
  • https://www.sesame.com/research/crossingtheuncannyvalleyof_voice#demo
  • https://github.com/SesameAILabs/csm

Platform and Feature Updates

GPT-4.5 Release and Integration: OpenAI has released GPT-4.5 with new capabilities including an effort level slider, and the model has already been integrated into Perplexity AI's platform.

• Which model will have an effort level slider? What kind of innovation are they trying to bring?
• GPT-4.5 is available on Perplexity now. THEY ARE FAST
• 4.5 triggered Deep Research to solve NYT word game
  • https://i.imgur.com/iatTViN.jpeg

Perplexity Deep Research for Enterprise Data: Perplexity AI has introduced a new enterprise feature that connects to Google Drive, OneDrive, and SharePoint, enabling comprehensive research across company files and the web.

• Introducing Perplexity Deep Research for Enterprise Data

AI Capabilities and Applications

Gemini 2.0 for Music Generation: A user implemented a new "reasoning" mode for music models using Gemini Pro 2.0, resulting in significant improvements in AI-generated music.

• Using Gemini 2.0 to implement a new reasoning or "thinking" mode for music models produces unbelievable results