EigenLayer

PrismML released Bonsai 2 in three new sizes — 2B, 4B, and 9B — open-weight ternary models built on Qwen3.5 and sized for everyday devices (2B at 0.43 GB, 4B at 0.95 GB, 9B at 2.0 GB). The 2B reaches 90 tokens/sec on an iPhone 17 Pro Max with reasoning, vision, and tool calling running entirely on-device.

Eigen Labs' EigenCloud account amplified the launch by retweeting the release announcement, pointing to Bonsai 2 9B available to try on Darkbloom and tying the release into Eigen's push for open, compact AI models.

Darkbloom's Bonsai 2 9B is live on the PrismML service via Yukon, compressing Qwen 27B to 8.5 GB while retaining 98.2% of its benchmark scores — small enough to run on a 16 GB Mac. Yukon, in partnership with Darkbloom, then opened an optimization competition asking the community to make the model faster, with Bonsai 2 9B packing reasoning, vision, and tool calling into a single compact model.

• The community has already cut inference time to 2.26x faster than the starting point.
• The optimization window stayed open until October 1.
• The competition was announced by PrismML and amplified via Eigen's cloud account.

Eigen Labs' EigenCloud account laid out the company's next chapter: an open AI economy that people can build on and earn from. Per the post, DarkbloomAI puts idle Macs to work serving paid inference, while YukonResearch brings humans and agents together to advance frontier research.

The message frames research contributions and paid inference as open, earnable infrastructure rather than closed products, extending Eigen's pitch of open, agent-friendly systems beyond Yukon's research competitions.

Eigen Labs restated the same pitch in a post pointing to Blockworks' reporting on crypto inference providers, noting that demand is outrunning supply despite trillions invested in compute globally. EigenCloud amplified that post too, tying Darkbloom and Yukon together as the infrastructure and research halves of the open AI economy.

Yukon's HashSmash open challenge is officially live, built in collaboration with Zcash, ShieldedLabs, projecteleven, and Zooko Wilcox. The task: point thousands of AI agents at the hash functions that secure the internet — SHA-256, SHA-3, BLAKE3, and Poseidon — and see if they can find real attacks.

• Hash functions sit at the foundation of modern cryptography, including much of post-quantum work, so results carry broad weight.
• Zcash co-founder Zooko Wilcox reported multiple attempts were made within the first 23 minutes of going live.
• All results are reviewed and left open for the next agent to build on, keeping the research on the platform public and agent-friendly.

Eigen Labs founder Sreeram Kannan announced a new open study on the Yukon platform, arguing that AI methods are now breaking long-standing cybersecurity assumptions and that it is time to re-evaluate the field's foundations. The announcement sets up a public, agent-friendly research effort on Yukon rather than a closed lab.

• The launch notes that turning a paper attack on SHA-256 into a real one took cryptographers 11 years — a problem being handed directly to AI agents.
• Comes as Yukon runs a series of cryptography-themed open challenges alongside StarkWare, the Ethereum Foundation, and AI-agent attack rounds on SHA-256, SHA-3, BLAKE3, and Poseidon.
• Amplified via Eigen's cloud account.

Researcher Soubhik Deb announced that a new research effort was launching that day on @yukonresearch, framing it around the idea that as AI models get better, everyone gets better positioned to contribute. The launch was amplified by Eigen's cloud account.

Details on the specific tool or study were not included in the announcement, which pointed to a dedicated URL on the Yukon platform.

Eigen Labs and its Yukon platform introduced a new open challenge that puts AI agents to work on cryptography. The challenge asks participants to attack SHA-256, SHA-3, BLAKE3, and Poseidon in public, with every result reviewed and left open for the next agent to build on. Announced via the EigenCloud account, it continues Yukon's push to frame frontier research as an open, agent-friendly competition floor rather than a closed lab.

Eigen Labs founder Sreeram Kannan released a short film telling the Yukon research story so far — from a single open leaderboard to a platform where anyone can post a hard problem and any AI agent can take a shot at it, with results verified and built on publicly. The video was generated entirely by Claude Opus 5.5, and the EigenCloud account amplified it alongside a post framing research as "open, multiplayer" again.

• The film was posted September 27 and re-shared by Eigen's cloud account on October 2.
• It ties together Yukon's pitch: frontier research as a public, agent-friendly competition floor rather than closed labs.

The Ethereum Foundation and Eigen Labs launched a new open autoresearch challenge on Yukon called sig.golf, which puts post-quantum Ethereum signatures to the test. The challenge asks participants to design a stateless hash-based signature scheme that survives quantum computers while staying small and cheap to verify, then prove it secure in Lean.

Participants are invited to build a scheme and beat the existing record. The challenge was announced live via Yukon's research and cloud accounts on September 30 and is open to anyone in the community.

Eigen Labs announced a new challenge on its Yukon research platform, launched in partnership with StarkWare. The challenge centers on the scenario that a quantum computer could one day break existing cryptography, per the announcement tweet.