Skip to main content
Back

Grok 2.5 open-sourced: full weights on Hugging Face

Meta's new reasoning method, GitHub ships MCP Server, Microsoft adds Excel Copilot, Meta x Midjourney, Anthropic's resea

August 25, 2025 3 min read

Hey

Welcome to AlphaSignal, the most read source of news by AI engineers and researchers. Every day, we identity and summarize the top 1% of news, papers, models, and repos, so you’re always up to date.


Here’s today’s roundup:

SUMMARY

Read time: 3 min 53 sec

Top News

▪ xAI open-sources Grok 2.5, its flagship reasoning model from last year

Intercom

▪ Join candid technical panels on architecture, scaling, and model training.

Trending Signals

▪ Meta publishes a method to boost model reasoning

▪ Anthropic explores filtering training data to lower future misuse risks

▪ GitHub releases open-source MCP Server, connecting AI tools directly to GitHub

▪ Microsoft adds COPILOT function, bringing AI prompts directly into Excel

▪ Meta partners with Midjourney to integrate aesthetic AI into future models

Top Repos

▪ googletest: Google’s framework for reliable C++ testing and mocks

▪ sim: Build and deploy AI agent workflows visually with LLM connections.

▪ parlant: Enforce guidelines in AI agents with tools and explainability

AI Coding Tip

▪ Use OpenAI’s GPT‑5 guide to refine your prompts for coding

If you're enjoying AlphaSignal please forward this email to a colleague. 

It helps us keep this content free.

TOP_NEWS

xAI open-sources Grok 2.5

50,375 Likes

alpha_signal_image_2

xAI released Grok 2.5, open-sourcing the 270B-parameter model that served as its flagship system in 2024. For the first time, you can download the full weights and run inference locally with deployment instructions for multi-GPU setups.


Open-sourcing gives developers the ability to study architectural trade-offs, measure real-world efficiency, and test interpretability challenges in modern MoE systems.


What This Tells Us

  • Grok 2.5’s shared expert design reduces inference costs by always keeping one expert active while routing others dynamically.

  • Compared to traditional MoE, the model is larger overall but activates only ~62B parameters per token, balancing scale with efficiency.

  • With clear setup details, the model can be used as a reference implementation for serving ultra-large models under GPU constraints.

How to Use It

  • Download the weights from Hugging Face (~500 GB, 42 files).

  • Install SGLang v0.5.1+ for inference.

  • Serve the model with tensor parallelism across 8 GPUs.

  • Use the provided chat template to run post-trained checkpoints correctly.

License and Restrictions

  • Free for research and commercial use, as long as you follow xAI’s Acceptable Use Policy.

  • You cannot use it to train or build other foundation models, except to fine-tune Grok 2 itself.

  • Redistribution must include the license, attribution to xAI, and display “Powered by xAI.”

  • Outputs from the model are unrestricted for downstream use.

Availability

  • Available now on Hugging Face (xai-org/grok-2).

Try now
alpha_signal_image_3

Join AI experts to learn how leading products are built

Hear from AI leaders on training models, designing custom architectures, and integrating with real-world systems.


Join open, technical conversations in San Francisco on September 11.

  • Learn how to design and deliver reliable frontier AI products

  • Hear candid lessons from leaders across the AI stack

  • See Fin AI Group’s custom retrieval and reranker models in practice

  • Gain practical insights from unscripted, technical discussions

Register now

partner with us

TOP_SIGNALS

Meta introduces DeepConf, a parallel thinking method that improves reasoning efficiency with only ~50 lines of code

2,639 Likes

Anthropic tests six classifier approaches, finding efficient setups for removing hazardous data during pretraining

1,938 Likes

GitHub unveils open-source MCP Server, giving agents direct structured access to GitHub data and actions

538 Likes

Microsoft ships COPILOT function in Excel, letting you to run AI prompts directly in spreadsheet cells, now available to Beta users

21,476 Likes

Meta teams with Midjourney in technical collaboration to enhance visual capabilities

5,639 Likes

TOP_REPOS

GoogleTest

36,937 Stars

This Google repo runs C++ tests automatically. It includes tools to fake functions using simple syntax, matchers, and partial mocks. It checks everything without needing you to record or replay test steps.

sim

12,148 Stars

An AI agent workflow builder. It offers a simple, visual interface to rapidly build and deploy LLM‑based workflows. It supports drag‑and‑drop logic, instant API deployment, and self‑hosting via NPM, Docker, or dev containers.

parlant

7,724 Stars

Python framework for building LLM-powered agents that strictly follow instructions. Define conversation journeys and behavioral rules in natural language. Attach tools, use response templates, and enable explainability.

AI_CODING_TIP

OpenAI’s Tips for Better Coding with GPT‑5

OpenAI published a cheat sheet to help you use GPT‑5 more effectively for coding. It focuses on writing better prompts, reducing confusion, and managing complexity in agentic workflows.


Key Tips:

  • Use XML-like <tags> to structure rules and examples.

  • Set reasoning_effort to high for complex tasks, low/medium for speed.

  • Break down prompts to avoid mixing unrelated goals.

  • Don’t use strong language like “be thorough” or “do not miss anything.”

  • Include a hidden rubric to guide planning and decision steps.

  • Use tool budgets, turn limits, or check-ins to keep agents on track.

Read now

At Alpha Signal, our mission is to build a sharp, engaged community focused on AI, machine learning, and cutting-edge language models, helping over 200,000 developers stay informed and ahead. We’re passionate about curating the best in AI, from top research and trending technical blogs to expert insights and tailored job opportunities. We keep you connected to the breakthroughs and discussions that matter, so you can stay in the loop without endless searching. We also work closely with partners who value the future of AI, including employers and advertisers who want to reach an audience as passionate about AI as we are.


Our partnerships are based on shared values of ethics, responsibility, and a commitment to building a better world through technology.Privacy is a priority at Alpha Signal. Our Privacy Policy clearly explains how we collect, store, and use your personal and non-personal information. By using our website, you accept these terms, which you can review on our website. This policy applies across all Alpha Signal pages, outlining your rights and how to contact us if you want to adjust the use of your information. We’re based in the United States. By using our site, you agree to be governed by U.S. laws.

Looking to promote your company, product, service, or event to 200,000+ AI developers? Let's work together.

Work with us