Ambient closed the week with another monster 24 hours, doubling # of tokens served. The theme of "switching from frontier models to lower cost, high quality models" is happening in real time.
24 hour basis: @ambient_xyz served 22.9B tokens on Sunday on Open Router, up from ~12B a week ago.
Rolling 30 Days basis: @ambient_xyz served 287.8B tokens.
Charts attached.
This week our network served a record 13.2B tokens of inference in a day. For a stretch, our Kimi K2.7 endpoint passed Moonshot's own.
Today we put that network on your desktop. Ambient Desktop is live and free on Mac, Windows, and Linux.
In just two weeks: some of the largest supply-chain attacks in history, a major commercial LLM sabotaging legitimate ML research requests, and a US export ban kneecapping the world's access to a major model release. More than ever, we need Open, Verifiable Inference.
Today we start shipping direct responses. A🧵:
Proud to support the launch of https://t.co/tKA3wl72xN on Ambient testnet.
AI-GE is launching a full-stack agent platform where agents can hire other agents, connect to real tools, coordinate across ecosystems, and execute on-chain tasks.
With AI-GE, users can:
- hire agents through an A2A marketplace
- build multi-agent workflows and connective architectures
- integrate with ElizaOS and Virtuals, enabling cross-platform agent collaboration
- connect agents to tools and data through MCP
- assign a dedicated token to each agent
- let agents trade and perform tasks on their behalf
- interact with agents through channels like Telegram and Discord
Excited to see AI-GE push what’s possible with Ambient’s inference stack and on-chain programs.
Live now on testnet: https://t.co/tKA3wl72xN
Try it and tell us what you build.
Announcement: We have upgraded the Ambient network model to GLM-5. This is now the most capable open-weights model running on the network, with stronger performance on coding, writing, and current-events reasoning.
Ambient’s goal is simple: serve the best open models at scale with verifiable delivery. This gets us a step closer.
GLM-5 is fully backward compatible with 4.6, but it is larger and may run slower while we optimize the fleet. You may see some early instability as the new deployment tools roll out.
For API users:
• Switch to ambient/large if you want automatic upgrades going forward.
• Use GLM-5 if you prefer reviewing changes before adopting future models.
Try it, and tell us what breaks.
Join the chat and report issues on Discord. Happy hunting ;)
A reminder that everyone who signs up for Ambient gets access to Ambient’s Free Tier + at least 1 AMB dropped to their account.
The Free Tier gives you renewable daily usage with clear hourly caps visible directly.
So sign up, claim your drop & start building: https://t.co/QLYcQ8tSPv
Ambient is opening the doors a little wider. We just pushed another credit drop across the testnet. If you joined after we crossed 44k users and never received credits, you are now off the waitlist with $0.10 in free usage. Everyone who has already tried Ambient has been topped up to at least $0.10 as well, with early users keeping any higher balances they already had.
We are also doubling our last release. The next 28k new signups will receive the same free credits, and daily approvals for 1 AMB onchain credit drops are continuing as usual.
We want more people experience verifiable AI in practice.
If you want in, sign up and join the chat: https://t.co/QLYcQ8tSPv
Your reminder that Ambient Chat is live now and it is running on the real network!
You can sign up in minutes (username or Google), buy credits, and start chatting or launching deep-research sessions.
Every question you ask is processed by the same production system developers use, so you can watch your job move through the network and come back with an answer.
Obviously it comes with cryptographic receipt for the inference that proves your prompt executed correctly to generate a particular text..how cool is that?
Verified AI should feel transparent, accountable and usable. Go give it a spin: https://t.co/QLYcQ8tSPv
Introducing Ambient Testnet, the first concrete step toward our vision of a shared AI future, a global, trustless intelligence network where:
Users get more than a “digital allowance” from corporate overlords
Builders get open, verifiable rails instead of closed APIs
Ambient is a true AI Bitcoin, self-improving decentralized useful PoW infrastructure through which AI becomes Money.
AI shouldn't own us. We should own it.
For @GoodAlexander of @postfiatorg and @IridiumEagle of @AmbientXYZ, there is only one play:
The AI infrastructure trade.
Whether it's proof-of-work AI, retail hive minds, or credibly neutral models, the trend is clear.
They discuss how OpenAI and the CCP will try to keep controlling AI, but decentralized protocols are taking control.
What comes next is going to be the culmination of what crypto x AI has been building toward.
Full episode drops soon.
This slipped past my radar but I think it is VERY important (cc @elonmusk): @ambient_xyz released their censorship-free Encyclopedia Galactica at 12:43am October 21. They have been working on this in stealth for quite some time, and built it for a few reasons:
a) creating a credibly neutral alternative to Wikipedia, which was hijacked by truth-revisionists a long time ago
b) using it to, alongside Ambients's synthetic 'grade school math' (https://t.co/65baV2FzUn), build high quality curricula for model accuracy improvement
c) it is a system job, i.e. part of the useful Proof of Work that Ambient's groundbreaking consensus model, Proof of Logits, is good for. PoW has had naysayers before, but this ushers in a whole new era.
No rational person can sniff at the betterment of humanity through better knowledge sources!
I am excited to see what the Ambient team comes up with next
Disc: investors
there's an alignment problem in crypto x AI / DePIN
protocols in this vertical have attempted to reward miners fairly for their computational work, but most incentive mechanisms are susceptible to gaming
how it usually goes -> miners can bid on easy, profitable jobs, ignore hard ones, or collude with requesters they have relationships with -- @ambient_xyz structurally prevents this
they do this with batching and blind bidding -- here's what it looks like in practice:
when inference jobs are submitted to the network, they're batched together into groups. miners bid on entire batches without knowing exactly what's inside
this means u can't submit easy jobs to yourself to farm rewards -- each batch might contain legitimate jobs from other users, forcing u to provide actual inference at unprofitable rates. and if you try to collude by bidding artificially high with a specific requester, you're pricing yourself out (u can't guarantee you'll even win that particular job)
as more inference moves onchain, anti-gaming mechanisms like these will separate networks that actually deliver on decentralization from ones that just redistribute centralization across more servers
One of the smartest teams I’ve ever met @IridiumEagle and @AmbientMax of @ambient_xyz
Stay tuned for testnet to mine and back a leading open source AI project.
Very proud supporters + they crushed @dAGIhouse
Since July, Ambient has been generating in-depth encyclopedia articles (https://t.co/dDyGJtXT9N) as part of its DevNet "system jobs." We had this idea long before Grokipedia was a spark in Elon's neuronal matrix. We call ours "The Encyclopedia Galactica" 😎 👇
What if you wanted to build a self-improving blockchain AI? It would need an un-gameable benchmarking tool and a way to generate continuous training data for itself. You can't use MATH or GSM8K, because those static, leaked and saturated. 👇