Reasoning Models Go Lean, Multimodal Local Inference Expands, and Open Agents Get Sharper Tools

Research

Open-R1: Fully Open Reproduction of DeepSeek-R1 Training Pipeline

The Open-R1 project publishes a complete replication of the DeepSeek-R1 reasoning training pipeline, giving the community a reproducible path to train reasoning models from scratch. This is the kind of infrastructure work that compresses the gap between frontier proprietary reasoning and what builders can actually run and fine-tune themselves.

Scaling Reasoning Without Scaling Parameters: Efficient Chain-of-Thought via Process Reward Models

This paper shows you can significantly boost reasoning quality on smaller models by training a process reward model to guide chain-of-thought, rather than just scaling up parameter count. Tactically useful if you want reasoning-quality outputs from a 7B or smaller model without paying the compute bill of a 70B.

Tools

Microsoft BitNet b1.58 2B4T: 2B Parameter Model Running on CPU at Practical Speeds

BitNet's latest 2B model runs inference on a standard CPU at speeds that are actually usable, using 1-bit weights that slash memory requirements. For anyone building on-device or edge deployments without GPU access, this is one of the more concrete proof points that CPU inference is not just a toy anymore.

Analysis

AgentBench 2.0: Updated Benchmark Shows Open Models Closing the Gap on Agentic Tasks

AgentBench 2.0 results show open models have meaningfully closed the gap on proprietary models across a range of agentic task types, with some open 32B models now matching GPT-4-class performance on tool use and multi-step reasoning. This is the clearest quantitative signal yet that open agentic stacks are production-viable.