Skip to main content
Back

Sarvam AI $234M raise , MiniMax M3 open weights , Claude Code 98% g

Sarvam raises $234M, MiniMax opens weights, Kimi K2.7 beats GPT-5 on math, Ng ships free local AI agent, transformer burned into custom chip

June 16, 2026 7 min read

Hey,

The real story today isn't any single headline. It's the same insight showing up everywhere at once: the model is not the product. The infrastructure around it is.

Claude Code is 98% guardrails. Hermes agents now carry a Stripe wallet, but your credentials never touch the conversation. Sarvam just raised $234M not to rent AI, but to own the whole stack.

Control is the new moat. Everyone's racing to build it.
In Partnership with · partner_image
Announcement

AI coding tools have made your team faster. So why does it feel like there are more fires to put out than ever?

The bottleneck didn't disappear, it just moved.

On June 25th at 10am PT, AlphaSignal is hosting a live webinar with Gal Vered, CEO of Checksum, who's spent years solving exactly this problem at scale.

We're getting into it: why AI-generated code keeps failing in production, what engineers are missing in their verification process, and what it actually takes to ship code you can trust.

Join us for this free webinar and stop guessing whether your AI-generated code is ready to ship.

Register Now
Summary

Read time: 7 min 8 sec

Top News

▸ Sarvam raises $234M Series B at $1.5B valuation to build sovereign AI

Unblocked

▸ [Webinar] 8 levels of context in AI engineering

Top Model

▸ MiniMax opens model weights to the public

AssemblyAI

▸ Your voice agent is only as good as its STT

Top Repo

▸ Nex catches Rio 3.5 secretly built on their open-source model without credit

Signals

1. Kimi K2.7 ranks 2nd on math benchmark, beating GPT-5 and Grok 4.3
2. Presented by Braintrust: Braintrust explains how to turn production traces into golden datasets using human review
3. Andrew Ng ships a free desktop AI agent that reads files and sends emails locally
4. Claude Code is 98% traditional software, not AI, study finds
5. Someone pruned Qwen3's 35B model down to 6B params and it fits in 3.4 GB
6. Developer burns a full transformer into a custom chip hitting 56,000 tokens per second
Top News
Sarvam raises $234M Series B at $1.5B valuation to build sovereign AI
8,463 Likes
Sarvam raises $234M Series B at $1.5B valuation to

Sarvam AI just hit unicorn status, raising $234M in its Series B at a $1.5B valuation. The big idea: countries and companies should own their AI, not rent it from foreign providers.

That's sovereign AI in plain terms. Instead of depending on models built abroad, you build and control your own. Sarvam is betting that's a massive opportunity.

Here's what the money goes toward:

  • Training the next frontier model, focused on agentic AI (AI that takes actions, not just answers questions), coding, and cybersecurity
  • Expanding compute infrastructure to handle more deployments at scale
  • Models that understand Indian languages, voices, and documents

They already have real traction. 10 million API calls per day, models digitizing 35 million pages of records, and deployments across insurance, land records, and agriculture.

HCLTech dropped $150M as lead investor. For you, this means a growing platform with actual API access, Indian-language models, and an expanding stack to build on.

Read more
forward →
Presented by Unblocked
[Webinar] 8 levels of context in AI engineering

AI shows up in 60% of engineering work. But only about a fifth of it can be handed off without someone babysitting the output. That’s because agents are missing context.

This 8-level context maturity model shows why your output is still inconsistent after all the tokens burned.

Join Unblocked live June 24 (FREE) to learn:

  • Why more MCPs provides agents access but not understanding
  • What it takes to deploy agents you can trust without supervision
  • How a context layer solves for quality, efficiency and cost
Register now
partner with us →
Top Model
MiniMax opens model weights to the public
6,523 Likes
MiniMax opens model weights to the public

The weights for MiniMax M3 just dropped on Hugging Face. Here is what you need to know.

M3 is the first open-weight model to combine frontier coding, a 1 million token context window, and native multimodal support (images and video) in one package. That combo was locked behind paid APIs until now.

The 1M context window is the big deal. A new sparse attention design drops the compute cost at 1M context to roughly 1/20th of the previous generation, so it does not cost a fortune to run long contexts.

On the coding side: M3 scores 59.0% on SWE-Bench Pro, ahead of GPT-5.5 at 58.6%.

  • Feed it an entire codebase at once, no chunking needed
  • Send images and video directly, no separate vision model
  • Control a desktop computer, similar to Anthropic's computer-use API
  • Toggle a thinking mode on or off per request

API pricing starts at $0.30 per million input tokens. Weights are now on Hugging Face to self-host.

Read more
forward →
Presented by AssemblyAI
Your voice agent is only as good as its STT

Most developers find their STT's limits in production, not before.

Universal-3 Pro delivers 94.2% accuracy on accented speech, medical terms, and proper nouns.The foundation your voice agent needs to ship.

Try it yourself
partner with us →
Top Repo
Nex catches Rio 3.5 secretly built on their open-source model without credit
5,338 Likes
Nex catches Rio 3.5 secretly built on their open-s

A city government in Rio de Janeiro released a model called Rio 3.5, claimed it was original, and the AI world went nuts. Then Nex AGI did the math.

Turns out Rio 3.5 is just a weight merge (think: blending two existing models together like mixing paint) of two open-source models: 60% Nex N2 Pro + 40% Qwen 3.5. No original training. No novel research. Just a blend with a new name.

The smoking gun? Strip away Rio's system prompt and the model literally introduces itself as Nex N2 Pro.

Here is what you actually get with the real Nex N2 Pro:

  • Free, Apache 2.0 license, weights on Hugging Face
  • 75.3 on Terminal-Bench 2.1, top-3 globally, beating Claude Opus 4.7
  • Adaptive Thinking: the model decides how deeply to reason per task, cutting wasted compute by up to 50%
  • Run it free today via OpenRouter or deploy locally with their sglang fork

The takeaway: open-source attribution matters. And the real model was free the whole time.

Read more
forward →
Signals
1
Kimi K2.7 ranks 2nd on math benchmark, beating GPT-5 and Grok 4.3
4,846 Likes
2
Braintrust explains how to turn production traces into golden datasets using human review
Presented by Braintrust
3
Andrew Ng ships a free desktop AI agent that reads files and sends emails locally
14,478 Stars
4
Claude Code is 98% traditional software, not AI, study finds
1,289 Likes
5
Someone pruned Qwen3's 35B model down to 6B params and it fits in 3.4 GB
572 Downloads
6
Developer burns a full transformer into a custom chip hitting 56,000 tokens per second
3,833 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