Skip to main content
Back

Meta's 816B Model Runs Locally on Two Mac Studios

Meta ships 816B coding model, GPT-6 solves 98% of expert physics problems, Google releases Gemini 4 Argon, Black Forest Labs drops 5.6B image model

October 16, 2026 6 min read

Hey,

Open weights keep getting heavier. Meta just dropped an 816B model you can run on two Mac Studios, while GPT-6 is correcting physicists' own homework at 98% accuracy.

The pattern is clear: frontier capability is becoming a commodity, fast. What took a GPU cluster now fits in your garage. What took a PhD now costs $3.

We're not just watching AI get smarter. We're watching the moat evaporate in real time.
In Partnership with · partner_image
Author photo
Today's Author
Lior Alexander. Founder of AlphaSignal and former ML Engineer. Previously built ML systems at Iguazio, Guesty, Enphase, Mila.
Summary

Read time: 6 min 41 sec

Top News

▸ Meta releases 816B open-weight model built for coding and long agentic tasks

Mistral

▸ Mistral Large 4 for Agentic Workflows

Top Repo

▸ Cheng Lou's open-source tool converts JavaScript code into a real neural net without training

Mercor

▸ Frontier Models Aced This Accounting Study

Top News

▸ OpenAI ships GPT-Live-1.1 with new voices in six languages to all users

Signals

1. Google releases Gemini 4 Argon, turning prompts into apps, slides, and 3D files
2. AWS : See how AWS and SANS Institute approach API and agentic AI risks with stronger governance, observability, and controls
3. Meta releases AudioCraft, an open-source PyTorch library for music and audio generation
4. GPT-6 solves 49 of 50 expert physics problems at 98% accuracy for $3 each
5. Black Forest Labs releases 5.6B open-weights model for text-to-image and multi-reference editing
6. How to run uncensored open-source AI models locally using Ollama
Top News
Meta releases 816B open-weight model built for coding and long agentic tasks
9,882 Likes
Meta releases 816B open-weight model built for cod

Meta just dropped Muse Spark Open, and it's a big one. It's a massive 816 billion parameter model, but here's the trick: only 42B parameters actually run per request. Think of it like a team of 256 specialists where only 8 show up per task. You get huge model quality without paying for all 816B every time.

What can you actually build with it?

  • Coding agents that handle long, multi-step tasks without losing track
  • Apps that read text, images, video, and audio all at once
  • Workflows using a 1M token context window, meaning you can feed it entire codebases in one shot

The wild part: you can run this locally on two Apple M5 Ultra Mac Studios using the MLX framework. No cloud required. You can also hit it via API on Groq, Together AI, OpenRouter, and others, or grab the weights on Hugging Face. Fine-tuning is supported through PyTorch. The license allows free use for many projects, but check the Muse Model License for commercial limits.

Read more
forward →
Presented by Mistral
Mistral Large 4 for Agentic Workflows

Mistral Large 4 is now 50% off API pricing. It has 1 trillion parameters, native multimodal support and zero data retention.

It’s our strongest model for agentic coding: 61.7% on DeepSWE v1.1, and ahead of leading open models on the Coding Agent Index.

Test it in your stack:

  • Fix real bugs and coding tasks from your backlog
  • Query your codebase, legal contracts, filings or reports
  • Parse charts, diagrams and technical drawings

Call mistral-large-4 from the API.

Start building
partner with us →
Top Repo
Cheng Lou's open-source tool converts JavaScript code into a real neural net without training
9,040 Likes
Cheng Lou's open-source tool converts JavaScript c

What if your regular JavaScript code was already a neural net? That's the idea behind Handwired.

Neural nets are usually trained on data, which makes them fuzzy and probabilistic. Handwired skips training entirely. Instead, it converts your if-else statements and loops directly into neural net weights, numbers that represent your program exactly, with zero guessing.

Why does this matter? Because now your code is fully differentiable, meaning you can apply machine learning techniques like loss functions and backpropagation (nudging numbers to improve a result) to actual working software.

  • Runs entirely in the browser, no GPU needed
  • Search 100k JSON files in real time using WebGPU matrix math
  • Photos that learn to rearrange themselves as you drag
  • Mix it with an LLM so it can follow both learned behavior and exact program logic

You can grab the source and tutorial on GitHub and play with the demos live in your browser right now.

Read more
forward →
Presented by Mercor
Frontier Models Aced This Accounting Study

Mercor built a human baseline study: 12 CPAs vs. AI models on four real accounting tasks. Accountants averaged ~37%.

Gemini 2.5 Pro landed below that; Opus 5 now hits 100%, while open-weight Qwen3.5-122B still trails.

Read the full study, analysis, and methodology.

See the data
partner with us →
Top News
OpenAI ships GPT-Live-1.1 with new voices in six languages to all users
8,941 Likes
OpenAI ships GPT-Live-1.1 with new voices in six l

OpenAI just updated its voice model. GPT-Live-1.1 is now live for all users across ChatGPT, Codex, and Dots.

The big deal? Voice now sounds actually local, not like a robot guessing at accents. It brings 13 new voices, two per language, in French, German, Hindi, Japanese, Korean, and Mexican Spanish, plus Bubbie in English.

There are also two model sizes now. Plus and above users get GPT-Live-1.1 by default, while others get the smaller Quip version.

What else changed:

  • More natural accents across languages, and the voice uses a wider range of short phrases when it is searching instead of repeating "checking..." on loop.
  • More natural handoffs when delegating tasks, and better steerability so the voice responds the way you actually want.
  • API access is coming, but OpenAI has not provided a date.

To try it now: open the ChatGPT app, tap the Voice icon, and pick Live mode from Settings then Voice.

Read more
forward →
Signals
1
Google releases Gemini 4 Argon, turning prompts into apps, slides, and 3D files
8,685 Likes
2
See how AWS and SANS Institute approach API and agentic AI risks with stronger governance, observability, and controls
Presented by AWS
3
Meta releases AudioCraft, an open-source PyTorch library for music and audio generation
23,715 Stars
4
GPT-6 solves 49 of 50 expert physics problems at 98% accuracy for $3 each
2,053 Likes
5
Black Forest Labs releases 5.6B open-weights model for text-to-image and multi-reference editing
3,241 Downloads
6
How to run uncensored open-source AI models locally using Ollama
2,027 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 300,000 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 300,000+ AI developers?
Work with us