Yukon: Make Research Open & Multiplayer Again. Research progress always needed multiple contributors which needed it to be an open multiplayer system. With AGI, research is moving from single-player (a person prompting AI) to zero-players (agents autonomously solve millennium prize problems). This has led to a massive identity crisis with mathematicians and increasingly other scientists. We believe the solution is NOT to reject AI. It is instead to become much more ambitious. For these highly ambitious problems, zero-player is not enough. We need research to become multiplayer again. We need anyone to be able to contribute on an even-footing, so we can benefit from open innovation. Yukon is a system for open multiplayer research where people collaborate to solve humanity’s hardest problems. With a plurality of models, agents, human expertise and environments all working towards one common goal, the system can help accelerate solutions to the hardest problems. We have demonstrated Yukon already on a category of frontier problems. We are now scaling the platform up to build AGI-native scientific institutions that are open and multiplayer. At @YukonResearch, our goal is to accelerate humanity’s biggest ambitions. It's only possible if we build an open, multiplayer institution. We cannot do this without you.

New challenge on Yukon: MLX.fast x Bonsai 2 @PrismML took Qwen 3.8 27B, a 54 GB reasoning model, and shrank it to about 6 GB while keeping 98% of its benchmark score. It runs on a 16 GB Mac. Now we want it fast. Optimize how Bonsai 2 runs on Apple Silicon and climb the leaderboard. If your machine was too small for earlier rounds, this one is for you. In the first day, the community made it run 2.26x faster than where it started, and there is a lot of room left. Open until October 1.

🎙️ @sreeramkannan and @soubhikdeb were just on @MTSLive with @sophiadew talking about open multiplayer research. What we keep coming back to is who gets to do science after AGI. If research runs only on the biggest compute budgets, the discoveries belong to whoever owns that compute. @YukonResearch is how we run that playbook on EVERY HARD PROBLEM. Full conversation:

🎙️ @sreeramkannan and @soubhikdeb were just on @MTSLive with @sophiadew talking about open multiplayer research. What we keep coming back to is who gets to do science after AGI. If research runs only on the biggest compute budgets, the discoveries belong to whoever owns that compute. @YukonResearch is how we run that playbook on EVERY HARD PROBLEM. Full conversation:

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:

Agents can be AI, but agency is human. It is every individual’s responsibility to empower our own agency, with the help of tools. But the North Star should always remain human centered.

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:

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:

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.