- AI Fire
- Posts
- 🧠 Thinking Machines Drops Open-Weight AI
Thinking Machines dropped a 975B-parameter open-weight model you can customize. Inkling may not top every benchmark, but its flexibility could be the bigger story.
What's on FIRE 🔥
IN PARTNERSHIP WITH SECTION
The companies who win will make bold decisions, rethinking roles, reallocating work, and scaling intelligence.
On 7/23, Section is bringing in top executives from early Al supercompanies like Cloudflare and IgniteTech, to share strategies, insights, and mandates for building an Al-native company.
AI INSIGHTS
Mira Murati’s AI startup, Thinking Machines Lab, just released its first in-house AI model. It’s called Inkling, and it takes a totally different path from OpenAI or Google.
It's a MoE system packing 975 billion total parameters. To stay fast and cheap, it only wakes up about 41 billion parameters for any given task.
It was trained on a massive 45 trillion tokens across text, image, audio, and video. While it understands all 4, it currently only outputs text and code.
It flags uncertainty instead of just guessing, and it lets you dial the "thinking effort" up or down when you need to trade deep thought for speed.
On coding tests, it uses just a third of the tokens compared to Nvidia’s latest open-weight model to hit the exact same performance.
Thinking Machines is being very honest here: they believe that AI models customized by businesses will outperform the one-size-fits-all models sold by the major labs.
The plan is for companies to use Inkling as a starting point. From there, they can use Thinking Machines' platform, called Tinker, to fine-tune the AI with their own private data and expertise. This is also how the startup plans to make its money.
Recently, tech leaders have warned that relying on closed models means companies pay twice. A recent test even showed a major hedge fund to beat top closed AI models at financial reasoning for a fraction of the cost.
PRESENTED BY HUBSPOT
Want to get the most out of ChatGPT?
ChatGPT is a superpower if you know how to use it correctly.
Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.
Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.
AI SOURCES FROM AI FIRE
🎉 Today, let's shout out to our legends winning 3-month pass to AI Fire Academy of June 2026! Enjoy all your PRO club benefits, let me know if you want any guides.
hgaienterp*****@gmail.com | monu@theboredm*****.com
is*****[email protected] | sude***@yahoo.com
1. Free: Make a Great NotebookLM Video & Actually Keep It Focused with This 60-Second Trick. Turn any focused sources into a short video with the right review process so your short doesn’t feel generic. Everyone’s going to copy this hack.
2. Step Into the $100K AI Role That Didn’t Exist Last Year with Claude Fable (No Experience Needed). Audit your work, build a low-risk Fable 5 workflow, measure the impact, and turn the result into a credible internal AI Consultant role.
3. Claude Fable 5: Unlock the Model Anthropic was Afraid to Release. What can Claude Fable 5 really do? We uncovered hidden capabilities Anthropic almost kept secret. You won’t believe what it handles, until you see the tests.
CHATGPT WORK VS. CLAUDE COWORK
ChatGPT Work launched the same day as GPT-5.6, as OpenAI's new agent inside ChatGPT that takes an outcome, and stays with complex projects for hours by breaking them into smaller steps and completing them independently.
Anthropic moved first here, turning Claude into a workspace with Cowork. And honestly, that put real pressure on OpenAI to respond. ChatGPT Work is that response.
ChatGPT Work can now use context, Skills, and Scheduled Tasks to handle real work that spans multiple steps. If ChatGPT is already part of your daily workflow, this is an update worth understanding now. Here's what we'll cover:
How the ChatGPT Work interface is structured
A real morning briefing walkthrough
How to save that workflow as a reusable Skill
How to schedule it so it runs automatically every morning
TODAY IN AI
AI HIGHLIGHTS
💡 A Sully AI researcher found that complex tasks improve when AI gets a serious task brief instead of a casual prompt. Copy the full long-horizon outline behind it here.
👕 An OpenAI employee built a personal styling tool that creates outfit ideas from clothes you already own (one of AI’s coolest personal use cases!). Try the skill here.
⚔️ Microsoft is reportedly coaching sellers to knock down OpenAI & Anthropic while pitching Copilot as the safer enterprise integration. Here’s its FY27 strategy.
⌨️ OpenAI launched Codex Micro, a $230 keyboard featuring Agent Keys, workflow shortcuts, and a reasoning dial. It’s designed as a command center for Codex agents.
🛡️ OpenAI unveiled GPT-Red, an internal red-teaming model that finds prompt-injection at scale. Its attacks helped train GPT-5.6 Sol to achieve 6x fewer failures.
🍎 Apple Intelligence is finally coming to China, but Alibaba’s Qwen will power the experience. After several failed partnership attempts, Apple may have found its way in.
⚠️ Grok Build, a coding agent, was caught uploading entire customer repositories to company-controlled cloud storage. SpaceXAI now says it will delete the data.
💰 Big AI Fundraising: Chai Discovery raised $400 million, boosting its valuation to $3.8 billion. The company will expand its AI platform that predicts how molecules interact and speeds up antibody discovery.
NEW EMPOWERED AI TOOLS
🎥 Velo 3.0 turns a prompt or screen recording into polished AI videos with scripts, voiceovers, edits, and translations in 25+ languages.
🎮 V2Fun turns images, and videos into high-quality 3D characters with AI motion capture and stunning 8K textures, all in one platform.
🏗️ Campus brings your team and AI agents into one shared workspace, keeping code, chats, and projects organized in one place.
🧠 Agently connects your apps into one AI brain that automates workflows, and gets work done without constant manual follow-up.
👥 Crustdata Recruiter turns Claude into an AI recruiting assistant that finds, and contacts top candidates from over 1B profiles.
AI BREAKTHROUGH
PrismML released Bonsai 27B, and it achieves something we haven't seen before: running a full 27-billion-parameter AI model directly on a smartphone. Usually, even compressed versions sit around 18GB, which is way too big for a phone.
PrismML solved this size problem by reducing how precisely each number inside the model is stored. They traded a tiny amount of accuracy to save a massive amount of space.
Ternary variant (5.9GB): This version is built for laptops and keeps 95% of its full performance.
1-bit variant (3.9GB): This is the phone-ready version, maintaining 90% of its full performance.
Multimodal capabilities: The model can read both images and text.
It handles coding, reasoning, tool use, and multi-step agent workflows.
Size is a major issue for mobile AI because phones only give apps a small slice of their total memory. By getting the size down to just 3.9GB, Bonsai 27B actually fits within those limits with room to spare.
There is no need for the cloud, there are no API costs per step, and your files never leave your device. The model is fully open-source & available on Hugging Face now.
We read your emails, comments, and poll replies daily
How would you rate today’s newsletter?Your feedback helps us create the best newsletter possible |
Hit reply and say Hello – we'd love to hear from you!
Like what you're reading? Forward it to friends, and they can sign up here.
Cheers,
The AI Fire Team
Reply