Skip to main content
Back

Z.ai GLM-5 744B model beats closed AI on engineering tasks

Gemini hits 90% IMO-ProofBench, xAI holds all-hands, Google AI studio gets Omnibar, Ai2's open platform for robotics.

February 12, 2026 4 min read

Hey,

Your daily briefing is ready. You can finally take a break from the AI firehose.

Our algos spent the night splitting signal from noise and pulled the top news, models, papers, and repos.

Here’s the must-read:

Together with Sonar
Sponsor Logo
Summary

Read time: 4 min 46 sec

Top News

▸ Z.ai launches an open-source model for long-horizon agent engineering tasks

Sonar

▸ Register for Sonar Summit to learn how to build better AI software

Top Repo

▸ Karpathy releases a dependency-free GPT that trains in ~200 lines of pure Python

Anyscale

▸ Join this webinar on Feb 19 to overcome memory limits in LLM training

Top Paper

▸ Google publishes papers detailing Gemini Deep Think discoveries in math

Signals

▸ xAI restructures into four core teams and plans lunar AI satellite factories
▸ Anthropic updates free plan with file tools, connectors, and conversation memory
▸ Google adds personalized dashboard and Omnibar to AI Studio
▸ Anthropic moves to prevent AI data centers from raising consumer electricity bills
▸ Ai2 introduces MolmoSpaces, an open platform for embodied AI research
Author photo
Today's Author
Lior Alexander. Founder of AlphaSignal and former ML Engineer. Previously built ML systems at Iguazio, Guesty, Enphase, Mila.
Top News
China's Z.ai releases GLM-5, an 744B open-weight model built for complex systems engineering and long-horizon agent tasks
5,218 Likes
Grok 4 Fast Benchmark

Z.ai launches GLM-5, a 744B-parameter open-weights model built for long-horizon agent tasks and complex engineering workflows.

It expands training data to 28.5T tokens and activates 40B parameters per step through a Mixture-of-Experts design, which means only part of the model runs for each token to reduce compute cost. On Humanity’s Last Exam with tools, it scores 50.4, placing it near leading closed models.

GLM-5 targets work that goes beyond chat. It runs multi-step plans, manages context, and executes coding tasks over extended sessions. On SWE-Bench Verified, it scores 77.8. On Vending Bench 2, a one-year business simulation, it ends with a $4,432 balance.

Key capabilities and access:

  • 744B total parameters, 40B active per token

  • DeepSeek Sparse Attention for long-context efficiency

  • MIT license with open weights

  • API access at $1 per million input tokens

  • Local deployment via vLLM or SGLang

  • Integration with Claude Code and OpenClaw

You can download weights from Hugging Face or ModelScope, call the API, or switch your agent config to “GLM-5.”

Try glm-5
Presented by Sonar
Sonar Summit: a global conversation about building better software in the AI era
alpha_signal_image_1

Join Sonar Summit on March 3rd, a global virtual event bringing together the brightest minds in software development.

In a world increasingly shaped by AI, it’s more crucial than ever to cut through the noise and amplify the ideas and practices that lead to truly good code.

Sonar created Sonar Summit to help you navigate the future with clarity and gain the knowledge needed to build better software, faster.

Register now
partner with us
Top Repo
Andrej Karpathy publishes a single-file GPT implementation that trains and runs inference without external libraries
13,148 Likes
Grok 4 Fast Benchmark

Andrej Karpathy strips GPT training down to its mathematical core. He publishes a dependency-free Python file that trains and runs a small GPT in about 200 lines, reduced from 243 by simplifying autograd.

The file shows the complete transformer algorithm using only basic math operations such as addition, multiplication, exponentials, and logarithms.

He removes tensor libraries and GPU acceleration to expose the mechanics. A tiny scalar autograd engine computes gradients.

Each operation defines its forward value and local derivative, and backward() chains gradients from the loss. Adam updates parameters. Cross-entropy computes prediction error.

You run it as a single Python script. You adjust model size and context length directly in code. It runs on CPU and scales only to small models, but it exposes every gradient path clearly for study and experimentation.

Try now
Presented by Anyscale
Scaling LLM Fine-Tuning with FSDP, DeepSpeed, and Ray
alpha_signal_image_2

Hitting memory limits when fine-tuning LLMs is a sign you're ready to scale.

This session covers distributed fine-tuning with FSDP, DeepSpeed, and Ray, the open-source framework behind Cursor, xAI, and Apple.

You'll walk through:

  • ZeRO stages and mixed precision config

  • Distributed training jobs with PyTorch

  • Checkpointing and failure recovery

Limited seats to keep it interactive.

Register now
partner with us
Top Paper
Google shows how Gemini Deep Think hits 90% IMO-ProofBench, enabling autonomous math research
1,932 Likes
Grok 4 Fast Benchmark

Gemini Deep Think no longer stops at Olympiad problems. After reaching Gold-level performance at the IMO and strong results in competitive programming, Google moves it into real research across math, physics, and computer science.

The shift matters because research problems do not come with clean answers. They require proof, refutation, and constant revision.

To handle this, Google builds a research agent called Aletheia on top of Deep Think. It generates candidate proofs, checks them with a natural-language verifier, revises weak steps, or restarts if the logic fails.

It can also say “I cannot solve this,” which prevents wasted effort. It searches literature to avoid fake citations and computational mistakes.

Key features

  • Generator-verifier-reviser loop

  • Web search and literature grounding

  • Failure detection

  • Inference-time scaling for stronger reasoning

Reported results

  • 90% on IMO-ProofBench Advanced

  • Solutions to open Erdős problems

  • Advances in optimization, ML theory, auction design, and cosmic string physics

Read the paper
Signals
1 xAI hosts its first all-hands after SpaceX merger, outlines lunar data center vision to scale next-generation AI systems 18,421 Likes
2 Anthropic brings subscription features including connectors and skills to free users 6,883 Likes
3 Google updates AI Studio with Omnibar for faster search and project management 2,282 Likes
4 Anthropic commits to cover consumer electricity price increases from AI data centers and pledges 100% payment 4,712 Likes
5 Ai2 launches MolmoSpaces, open embodied AI platform with and a benchmark to evaluate generalist robot policies 924 Likes
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 250,000+ AI developers?
Work with us