• AI Dev and Research News
  • Posts
  • ☑ 10 Mins AI Read: Microsoft AI Released Phi-4-Reasoning and DeepSeek Released DeepSeek-Prover-V2

☑ 10 Mins AI Read: Microsoft AI Released Phi-4-Reasoning and DeepSeek Released DeepSeek-Prover-V2

Hi There,

Dive into the hottest AI breakthroughs of the week—handpicked just for you!

Top 5 AI News 🔥 

Sponsored

🧵 Diagnosing and Self- Correcting LLM Agent Failures: A Technical Deep Dive into τ-Bench Findings with Atla’s EvalToolbox

⇧2,900 Likes

OpenSource

⇧2,800 Likes

LLM & Reasoning

🧵 Microsoft AI Released Phi-4-Reasoning: A 14B Parameter Open-Weight Reasoning Model that Achieves Strong Performance on Complex Reasoning Tasks

⇧2,650 Likes

AI App

🧵 Meta AI Introduces First Version of Its Llama 4-Powered AI App: A Standalone AI Assistant to Rival ChatGPT. [Download the app]

⇧2,354 Likes

Open Source AI

🧵 DeepSeek-AI Released DeepSeek-Prover-V2: An Open-Source Large Language Model Designed for Formal Theorem, Proving through Subgoal Decomposition and Reinforcement Learning

⇧2,100 Likes

Coding Agents

🧵 Xiaomi introduced MiMo-7B: A Compact Language Model that Outperforms Larger Models in Mathematical and Code Reasoning through Rigorous Pre-Training and Reinforcement Learning

⇧1,800 Likes

Sponsored

🧵 Meet Parlant: The Fully Open-Sourced Conversation Modeling Engine

⇧1,500 Likes

Meta AI

Meta AI Introduces ReasonIR-8B: A Reasoning-Focused Retriever Optimized for Efficiency and RAG Performance

Meta AI has released ReasonIR-8B, a retriever model designed explicitly for reasoning-intensive information retrieval. Trained from LLaMA3.1-8B, the model establishes new performance standards on the BRIGHT benchmark, achieving a normalized Discounted Cumulative Gain (nDCG@10) of 36.9 when used with a lightweight Qwen2.5 reranker. Notably, it surpasses leading reranking models such as Rank1-32B while offering 200× lower inference-time compute, making it significantly more practical for scaled RAG applications.

ReasonIR-8B is trained using a novel data generation pipeline, ReasonIR-SYNTHESIZER, which constructs synthetic queries and document pairs that mirror the challenges posed by real-world reasoning tasks. The model is released open-source on Hugging Face, along with training code and synthetic data tools, enabling further research and reproducibility.

⇧ 1,449 Likes

LLM and Reasoning

DeepSeek-AI Released DeepSeek-Prover-V2: An Open-Source Large Language Model Designed for Formal Theorem, Proving through Subgoal Decomposition and Reinforcement Learning

A team of researchers from DeepSeek-AI has introduced a new model, DeepSeek-Prover-V2, designed to generate formal mathematical proofs by leveraging subgoal decomposition and reinforcement learning. The core of their approach utilizes DeepSeek-V3 to break down a complex theorem into manageable subgoals, each of which is translated into a “have” statement in Lean 4 with a placeholder indicating that the proof is incomplete. These subgoals are then passed to a 7B-sized prover model that completes each proof step. Once all steps are resolved, they are synthesized into a complete Lean proof and paired with the original natural language reasoning generated by DeepSeek-V3. This forms a rich cold-start dataset for reinforcement learning. Importantly, the model’s training is entirely bootstrapped from synthetic data, with no human-annotated proof steps used.

⇧ 1,449 Likes

Top 5 AI Coding Tutorials </>

🖥️ Building a REACT-Style Agent Using Fireworks AI with LangChain that Fetches Data, Generates BigQuery SQL, and Maintains Conversational Memory

🖥️ How to Create a Custom Model Context Protocol (MCP) Client Using Gemini

🖥️ Tutorial on Seamlessly Accessing Any LinkedIn Profile with exa-mcp-server and Claude Desktop Using the Model Context Protocol MCP

🖥️ A Coding Guide to Different Function Calling Methods to Create Real-Time, Tool-Enabled Conversational AI Agents

Top 5 Trending AI Guides/Reports 📖

How was today’s email?

At Marktechpost AI Media Inc, we connect over 1 million monthly readers and 30,000+ newsletter subscribers with the latest in AI, machine learning, and breakthrough research. Our mission is to keep the global AI community informed and inspired—through expert insights, open-source innovations, and technical deep dives.

We partner with companies shaping the future of AI, offering ethical, high-impact exposure to a deeply engaged audience. Some content may be sponsored, and we always clearly disclose these partnerships to maintain transparency with our readers. We’re based in the U.S., and our Privacy Policy outlines how we handle data responsibly and with care.

Looking to promote your company, product, service, or event to 1 Million+ AI developers and Researchers? Let's work together.