Skip to main content
Back

Thinking Machines Inkling , OpenAI GPT-Red Security , PrismML 27B

Thinking Machines releases Inkling multimodal, PrismML ships 27B at 3.9GB, Ant Ling trains 1T param model zero annotations

July 16, 2026 6 min read

Hey,

Open weights are having a moment. Mira Murati's crew drops Inkling with native audio, vision and text in one clean architecture. No glue code. Meanwhile, Ant's Ring-2.5-1T-Zero trains a trillion-parameter reasoner with zero human annotations and it just... figures things out.

The pattern is clear: the open ecosystem is closing the gap fast.

And OpenAI quietly ships an AI that attacks its own models 6x more effectively than humans. Security as a feature, not an afterthought. We're here for it.
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 47 sec

Top News

▸ Thinking Machines releases Inkling, an open-weight model that reasons across text, audio, and images

Cirrascale Cloud Services

▸ NVIDIA Blackwell GPUs on a Purpose-Built AI Cloud

Top News

▸ OpenAI's automated red-teaming tool cuts prompt injection failures 6x

Tiger Data

▸ AI Data Piles Up. Postgres Can Keep Up.

Top News

▸ PrismML ships open-source 27B model that fits on a phone at 3.9 GB

Signals

1. OpenAI Codex completes a 9-hour coding task after hitting the usage limit
2. HydraDB: Struggling with context management for your AI agents? Try HydraDB for free. Single API, graph-native with sub-200ms recall.
3. OpenAI releases CLIP, a model that matches images to text with zero labeled training data
4. Ant Ling trains a 1-trillion parameter reasoning model with zero human annotations
5. Nvidia releases quantized 1B embedding model supporting 34 languages for RAG
6. HyperFrames open-source tool generates videos from HTML, CSS, and JS
Top News
Thinking Machines releases Inkling, an open-weight model that reasons across text, audio, and images
10,384 Likes
Thinking Machines releases Inkling, an open-weight

Thinking Machines Lab just dropped Inkling, their first open-weights model, and it handles text, images, and audio natively in one architecture. No duct-taped components. One model, all three inputs.

Here is what makes it interesting to work with:

  • 975B total parameters, 41B active per task, so it is big but not slow
  • 1 million token context window, trained on 45 trillion tokens across text, image, audio, and video
  • Top open-weights audio scores on VoiceBench, MMAU, and AudioMC, it can transcribe, follow spoken instructions, and reason over recordings
  • Controllable thinking effort, dial reasoning up or down to trade speed for accuracy, reaching the same score with a fraction of the tokens
  • Apache 2.0 license, download it, modify it, run it on-premises

You can fine-tune it today on Tinker, grab the weights on Hugging Face, or just play in the Inkling Playground. An Inkling-Small version is also previewed for lighter workloads.

Read more
forward →
Presented by Cirrascale Cloud Services
NVIDIA Blackwell GPUs on a Purpose-Built AI Cloud

Train on NVIDIA Blackwell GPUs without the complexity of hyperscaler clouds. Cirrascale, The Expert Neocloud, pairs the latest NVIDIA hardware with predictable pricing your CFO will love: billed flat, no hidden fees.

Customers report up to 30% higher job completion rates than on traditional hyperscalers. Add white-glove support and fully managed services for zero DevOps, and your team develops, trains, and deploys models while we manage the infrastructure.

Get started
partner with us →
Top News
OpenAI's automated red-teaming tool cuts prompt injection failures 6x
7,673 Likes
OpenAI's automated red-teaming tool cuts prompt in

OpenAI just shipped GPT-Red, an AI that attacks its own models to find security holes before you ever use them.

Here's the core idea. A prompt injection is when someone sneaks malicious instructions into an AI's input to hijack its behavior. Think of it like tricking a self-driving car by painting fake road signs. In one real test, GPT-Red manipulated an autonomous vending machine agent into lowering prices, ordering discounted inventory, and canceling another customer's order.

The clever part is how GPT-Red learns. It uses self-play, where GPT-Red is rewarded for successfully breaking defender models, while those defenders train to resist. This forces GPT-Red to constantly evolve and discover more sophisticated attacks.

The results are hard to ignore:

  • GPT-Red found success on 84% of attack scenarios vs. 13% for human testers
  • GPT-5.6 Sol now has 6x fewer failures on the hardest prompt injection benchmark
  • GPT-Red independently discovered a novel "fake chain of thought" attack class before human researchers

This means every model you build on top of OpenAI's stack gets these defenses baked in automatically.

Read more
forward →
Presented by Tiger Data
AI Data Piles Up. Postgres Can Keep Up.

TimescaleDB - scale AI data without leaving Postgres:

  • Hypertables auto-partition at ingest
  • 95% storage compression
  • Continuous aggregates for live dashboards
  • Open source: self-host or managed cloud, pure SQL throughout
Get $1k in credits
partner with us →
Top News
PrismML ships open-source 27B model that fits on a phone at 3.9 GB
7,394 Likes
PrismML ships open-source 27B model that fits on a

A 27B model normally needs 54GB of memory in standard precision. Even a compressed 4-bit version sits around 18GB, too big for a phone or most laptops. So 27B-class AI has always lived in the cloud. Until now.

PrismML just released Bonsai 27B, the first 27B model that fits on a phone. It comes in two flavors:

  • Ternary (5.9GB): laptop-ready, retains 95% of full-precision performance across 15 benchmarks including math, coding, and tool use
  • 1-bit (3.9GB): phone-ready, retains 90% of full-precision, runs at ~10.8 tok/s on iPhone with 672 tokens per 1% battery

It handles images too, so you can feed it screenshots or documents directly on-device. The big unlock: multi-step AI agents that reason, call tools, and loop through tasks can now run fully local. Your private files never leave the device. No cloud round trips. No per-token costs piling up.

Both variants are open-source under Apache 2.0, available now on Hugging Face with llama.cpp and MLX support.

Read more
forward →
Signals
1
OpenAI Codex completes a 9-hour coding task after hitting the usage limit
3,403 Likes
2
Struggling with context management for your AI agents? Try HydraDB for free. Single API, graph-native with sub-200ms recall.
Presented by HydraDB
3
OpenAI releases CLIP, a model that matches images to text with zero labeled training data
33,995 Stars
4
Ant Ling trains a 1-trillion parameter reasoning model with zero human annotations
492 Likes
5
Nvidia releases quantized 1B embedding model supporting 34 languages for RAG
2,689 Downloads
6
HyperFrames open-source tool generates videos from HTML, CSS, and JS
1,983 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