Introducing Contrastive Language Model (CLM): an ultra-fast System One Model trained with a contrastive learning objective that connects states and actions.
CLM-8B is pre-trained on internet-scale data and delivers up to 9× faster inference than Jev ⚡ while achieving comparable performance across computer-use, gaming, and tool-calling tasks.
With lightweight fine-tuning, CLM-8B sets a new SOTA on challenging agentic coding benchmarks, such as DeepSWE (81.6%) and Terminal-Bench 2.1 (87.6%). In contrast, Jev fails to serve as an effective verifier for these long-horizon tasks.
We also build an efficient training and serving infra for CLMs by disaggregating states and actions, allowing their embeddings to be cached and reused independently. This substantially reduces inference latency in settings where the state evolves continuously while the action set remains fixed.
Finally, we establish scaling laws for CLMs and show that the test contrastive loss decreases predictably as a power law in training compute, model size, and dataset size.
📄 Blog: https://t.co/zwi9JOHKGx
💻 Code: https://t.co/rsHRYCGR8I
🗣️ Discord: https://t.co/Uqtdefvo3J
🤗 Data & Models: https://t.co/wdSWGGO3hu
More details on CLM’s architecture, data recipe, and scaling laws in the thread below 🧵
Introducing Bespoke Nimble: an open data, open model, open recipe for an open Jev.
Code and info: https://t.co/aC7kPejrcj
Model: https://t.co/snwKGdhn1I
Data:
* A new data curation recipe called contrastive data curation.
* Slightly change facts to generate negative data. This pushes the model to discriminate better and become a better decision maker. The calibration is implicit.
* Didn't do ablations but I think this is a critical piece!
* This also means training data doesn't need probabilities.
* Data covered 10 categories, and is fully synthetic.
* This data is split into train and eval.
Training
* LoRA finetune of Qwen3.5-9B.
* Distillation-free: we use Jev to only evaluate.
* No RL yet!
Serving
* Parallel constrained decoding as suggested by @NielsRogge and @harshagundal.
Results:
* The post-trained Qwen (Nimble) became substantially better on our curated eval: 66% for Qwen to 90% for Nimble. Jev is at 93%.
* 100ms on H100 and free to use on your macbook! Feel the AGI for free.
* 2 days of building in public. :)
Big caveat is that there is no standard benchmark to measure performance, and it's possible Nimble is much worse on other benchmarks compared to Jev. But it should be better than Qwen!
We thank @typesafeai for making Jev and the inspiring discussions in the community. Hope this release lifts all the boats and encourages more research and activity in this space.
Announcing jina-ocr-v1, our new visual document parser with 3.4B total parameters and 570M active parameters, with speculative decoding built in. Throw PDFs, scans, tables, charts, or invoices at it and get clean markdown back. Available on 🤗 & Jina Reader `x-respond-with` today
Reef is taking off! 🔥 Within 2 days since we open sourced it, Reef got ~300 GitHub ⭐️!
Reef is the first open-source infrastructure designed to evolve both model weights and the agent harness from live experience.
Reef already supports:
🧠Model evolution: SAO, TTT-Discover, OpenClaw-RL, with more recipes coming.
🛠️Harness evolution: SkillClaw, Meta-Harness, GEPA, with native integration for pi, OpenCode, Hermes Agent, and more harnesses coming.
Come and build Reef with us at https://t.co/IsAcJZFNQq !
Omarchy is blowing up. I've never been involved with anything in my career that has grown this fast. With Ruby on Rails, we had years to build solid institutions, teams, and relationships. With Omarchy, we've been forced to figure it all out in twelve days. It's exhausting, but also incredibly exciting.
It was never sustainable with just Ryan and me running everything. That's how it was, more or less, up until Quattro. Lots of other contributors, but all the responsibility was on us to make sure the ship stayed afloat, the servers didn't crash, and fixes got pushed out quickly.
Now it's time to build a proper institution. Durable, resilient, and competent. That's what we're doing here on Basecamp now. I'm spinning up teams for every facet of responsible distro management, and I'm getting an absolute outpouring of interest for all of them. Everyone wants to be part of this. We're winning hearts, minds, and volunteers at an astounding rate. Great! We need all of it to succeed.
Because make no mistake: There are many people who'd love to see this rocket blow up before it reaches the moon. Aggrieved Linux users who don't like the sudden attention their exclusive hobby has received. Competing Linux distributions that are seeing our numbers explode. Mac stans who've sunk their identity into an apple. And, of course, any of the haters I've picked up in my quarter-century career speaking bluntly on the internet.
They're not going to succeed. Because we've already become unstoppable. There's too much momentum, too much money, and too much support now backing this effort. We're living the Mandate From Heaven meme at the moment. And we are here to fulfill the prophecy: The Year of Linux on the Desktop!
That has been a joke for two decades. But by the end of the year, nobody at Apple or Microsoft is going to be laughing. They're going to be scrambling. Because neither of these proud organizations currently has any method to counter the speed, vision, or ambition with which we're going to accelerate into the future of personal computing.
This is the moment. This is the opening. This is our chance.
For thirty years, we've been subject to one OS overlord or another. Dictating how we compute. Choking off competitors through platform malfeasance. Tollboothing the distribution. That ends now. Because Linux is going to win. And Linux is free. As in beer, speech, and source.
But just because it's inevitable doesn't mean it's going to be easy. We have a lot of work in front of us if we actually want to make our mark. But there's never been a better time for this kind of delusional ambition. The age of agents is the unlocking factor. It sounds like a LinkedIn slogan, but it's true. Where the application of tokens goes, the innovation follows.
We can fix everything.
Let's do it together.
Let's go.
---
This is what I sent to the dozens of new volunteers who've signed up for teams within the new Omarchy organization yesterday. But we might as well broadcast our mission and intentions to the world too.
AI agents have gotten genuinely good at the work itself. The harder problem: getting multiple agents to work well with a team — and with each other.
Today we're introducing AgentConnect: the open-source, self-hosted, multi-agent alternative to Claude Tag. Teams and agents working together across Slack, Telegram, Discord, GitHub, and GitLab.
https://t.co/qf2pl4u3EY
We’ve decided to open-source a multi-agent harness we use internally at YC.
We call it “QM” and it’s meant to be easy to customize, like Hermes or OpenClaw, but useful for a whole company. We use it across accounting, legal, events, and engineering (including building QM itself!).
The whole project is under an MIT license. It is cloud-first and has Slack and web UI natively.