Darkbloom

PrismML @PrismML · Oct 15

Today, we’re releasing Bonsai 2 2B, 4B, and 9B.

Based on Qwen3.5, these new ternary models bring stronger reasoning, visual understanding, and tool calling to compact footprints designed for everyday devices:

• 2B: 0.43GB
• 4B: 0.95GB
• 9B: 2.0GB

For this release, we focused on what makes small models useful as local assistants: following instructions, avoiding repetitive loops, calling tools on mobile devices, and giving useful answers without unnecessary output. The 9B model uses roughly 40% fewer tokens on standard instruction-following and everyday math tasks. Solving the Riemann hypothesis can wait.

These models are successors to our first-generation Bonsai 1.7B, 4B, and 8B, and provide a foundation for bringing practical agentic functionality to local applications on personal computers, phones, and wearables.

The Bonsai 2 family now includes four sizes spanning 2B to 27B parameters. Developers can choose the balance of capability, memory use, and responsiveness that fits their application and hardware.

Available today under Apache 2.0.

View quoted post on X

Gajesh Verified account @gajesh · Sep 18

Darkbloom is the first model provider to support Ternary Bonsai 2 27B -- concentrated intelligence that fits on your phone.

Try it now: console.darkbloom.dev/chat

First 250 users get 100 million free tokens;

PrismML's new flagship model:

- a ternary compression of Qwen3.8 27B at 2bits per parameter.
- 8.5 GB total, 5x smaller than original
- keeps 98.2% of the Qwen's FP16 benchmark performance.
- 75% cheaper than Qwen 27B.

Qwen3.8 27B already operates comparably with Opus 4.6 and 5.6 Luna on certain tasks. This one does it in the memory of a phone. 262K context, image input, Apache 2.0.

From our first run on the network, on a single M5 Max with no caching:

- 35 tok/s decode at 1K context,
- 31 tok/s at 10K,
- 19 tok/s at 50K.

But we expect more performance gain coming in a few weeks! That's a full 27B reasoning model running comfortably on any Mac.

1,000+ Macs are serving on Darkbloom right now. Go try it out!!

Thank you to @BabakHassibi @SahinLale @HessianFree @rsadri_ml @tushar_bans @evaninwords and the whole PrismML team. This is exactly the kind of model Darkbloom was built for.

You can read the essay by Bonsai on Why Local AI Matters:

View quoted post on X

Kydo Verified account @0xkydo · Sep 18

Today's my birthday, and I could not have asked for a better gift.

Bonsai 2 fits on most phones shipping today (anything >8GB). And it outperforms Opus 4.6 and 5.6 Luna -- on tasks that would have sounded absurd to attempt anywhere two years ago.

I try to be disciplined about timelines. My most optimistic estimate for something like this was early-2027. It's September 2026.

Huge thank you to the @prismml team -- @BabakHassibi @SahinLale @HessianFree @rsadri_ml @tushar_bans @evaninwords -- for putting this into the world.

We're bringing Bonsai 2 to @DarkbloomAI tonight. The 1,000+ providers can test it right away. And whenever those machines aren't in use, they'll serve Bonsai 2 to anyone who wants to try it.

So few people set up a local model themselves, and a model this good shouldn't be gated by that. More tomorrow.

View quoted post on X

Gajesh Verified account @gajesh · Aug 27

Yayy!! Thank you, @MatthewBerman for covering Darkbloom.

Really great to see this; your optimistic energy is infectious and feels great to see more believers in our mission to keep accelerating the frontier of open compute networks.

Darkbloom is one shot at making anyone participate in this rapidly growing AI compute economy.

We're early, and moving as fast as we can. If you saw Matt's video (highly recc absorbing his amazing energy) and are passionate about this mission, we're hiring engineers towards improving the inference and adding new paths such as sandboxes. Reach out to me or @0xkydo 🏎️

Open Networks will accelerate Open Intelligence.

View quoted post on X

Kydo Verified account @0xkydo · Aug 27

MLX.fast turns one month old tomorrow 🎊🥳🎆 As an early celebration, we're kicking off our Gemma 4 challenge today.

Gemma 4 runs on Darkbloom today (six billion tokens served yesterday). Any speedup we find here shows up as extra throughput on machines that are already plugged in and serving.

I think this is the start of something special, and before we jump in I want to retrace our footsteps a bit and share why the Gemma challenge is more important than it might seem.

One month ago we launched our first challenge with @poolsideai to see how far we could push Laguna performance on Mac machines. It started as a DM between @gajesh and @eisokant and turned into a real thing almost immediately. MLX.fast wouldn't exist if @eisokant and the @poolsideai team hadn't taken a bet on us 🙏🙏

The results blew our minds. 2x on Laguna XS 2.1, and decode speed increased from 70 to 200+. More than the optimization itself, I am more impressed by the many contributors who came from across the world to participate.

But before we had time to process what that meant, Qwen 3.8 27B dropped and stunned the community with its size and performance, 100x smaller but performing comparable to Opus 4.6.

So we rolled up our sleeves and moved fast, migrating everything we'd just finished on Laguna over to Qwen.

Everyone took what they'd learned from Laguna and, with MTP support on day one, the leaderboard got to a 3x improvement in two weeks.

Around the same time, Darkbloom (@darkbloomai) hit a milestone of its own. We moved onto the paid tier on OpenRouter🙏which gave us a lot more room to scale the models we support and onboard more machines. Not one of those machines belongs to us.

That's when an obvious question showed up.

Why don't we upstream everything the community had done with Qwen into the Darkbloom production environment?

The machines already serving Qwen could serve it faster. Within a day, production traffic on Qwen was 2x faster, entirely off the back of work that happened in the open.

Which made me think we might be stumbling onto something here. What if MLX.fast can also function as the open R&D arm for Darkbloom?

People serve compute on Darkbloom, then rally in MLX.fast to make that compute faster, and every improvement flows back to the machines doing the serving. While everything is in the open.

The Gemma challenge is a test of this hypothesis. Gemma is the most used, most served model on Darkbloom right now. More than six billion tokens served yesterday alone, on machines that belong to people in this community. If that loop works anywhere, it should work here. And if it does, the incentives between Darkbloom and MLX.fast connect directly, and both platforms get more sustainable.

Going into this challenge, I'm a bit nervous. MLX.fast is only one month old. I don't know how much the two communities actually overlap (I think they do!). I don't know how folks will feel about working on a model that isn't at the frontier. But I'm excited to find out, and to keep learning alongside everyone who's helped us get here.

PS: Frontier model optimization isn't going away on MLX.fast. We've got something special coming later next week. I think you're all going to love it.

View quoted post on X