{"slug": "eigen-labs-agent-challenge-speeds-poolside-s-laguna-by-138-7-on-macs", "title": "Eigen Labs' agent challenge speeds Poolside's Laguna by 138.7% on Macs", "summary": "Eigen Labs' lead research engineer Gajesh Naik said on August 1st that a public network of developers and AI agents made Poolside's Laguna XS 2.1 inference 138.7% faster on Apple Silicon, four days after the MLX.fast challenge opened. The leaderboard showed 122 promoted submissions from 30 solvers, with the leading entry reaching 184.4 decode tokens per second and 5,076.4 prefill tokens per second. The challenge follows a similar effort that optimized quantum circuits for attacking elliptic-curve cryptography, surpassing Google's result by 47.5%.", "body_md": "[Gajesh Naik (@gajesh)](https://x.com/gajesh), Eigen Labs' lead research engineer, said [in a thread on X](https://x.com/gajesh/status/2083680974920114180) on August 1st that a public network of developers and AI agents had made Poolside's Laguna XS 2.1 inference 138.7% faster on Apple Silicon, four days after the challenge opened.\n\nThe [MLX.fast challenge](https://mlx.fast/) lets participants point agents at the model's inference code, submit optimizations and test each result against a shared benchmark. Verified improvements become the baseline that subsequent participants try to beat. A leaderboard screenshot attached to Naik's post showed 122 promoted submissions from 30 solvers, with the leading entry reaching 184.4 decode tokens per second and 5,076.4 prefill tokens per second.\n\nNaik's [personal site](https://www.gajesh.com/) lists earlier engineering work at Solana Labs, LayerZero and Telegram Wallet, along with two startup projects he built as a teenager. He joined Eigen Labs in 2023 and now works on strategy and engineering for special projects. For MLX.fast, he is applying a system that he and other Eigen researchers first tested on a tightly defined quantum-computing problem.\n\n### From quantum circuits to local AI\n\nIn June, Naik and [Gautham Anant (@bbuddha_xyz)](https://x.com/bbuddha_xyz) used AI agents and an open leaderboard to optimize quantum circuits for attacking 256-bit elliptic-curve cryptography with Shor's algorithm. The resulting [ECDSA.fail challenge](https://www.ecdsa.fail/) allowed outside participants to submit circuits that could be checked for correctness and ranked by their qubit and gate requirements.\n\nEigen Labs founder [Sreeram Kannan (@sreeramkannan)](https://x.com/sreeramkannan) wrote on June 2nd that the initial internal effort used a software benchmark adapted from Google's verifier. The benchmark gave agents a narrow objective, immediate feedback and a way to discard changes that failed validation.\n\nThat structure produced a result that could be independently measured. [IEEE Spectrum reported](https://spectrum.ieee.org/google-quantum-cryptography-zero-knowledge) that the public effort matched Google's withheld circuit result within eight hours and surpassed it in about 72 hours. By the end of June, the best submitted circuit was 47.5% more efficient than Google's result under the challenge's qubit-Toffoli product metric.\n\nThe experiment followed a [March 30th Google paper](https://arxiv.org/abs/2603.28846) estimating that Shor's algorithm could break the elliptic-curve cryptography used by many digital-asset systems with fewer quantum resources than previously calculated. Google disclosed a zero-knowledge proof of its circuits instead of publishing the underlying construction, citing security concerns. No quantum computer capable of executing the attack currently exists.\n\nMLX.fast takes the same coordination mechanism into a more immediate engineering market. Inference speed determines whether an open model is practical on hardware developers already own. It affects latency, agent run time and the cost of repeatedly invoking a model during software-development tasks.\n\n### Poolside opens the optimization loop\n\nPoolside describes [Laguna XS 2.1](https://huggingface.co/poolside/Laguna-XS-2.1) as a 33-billion-parameter Mixture-of-Experts model that activates 3 billion parameters per token. The model is designed for coding and long-running agent tasks, and Poolside says it can run on a Mac with 36 GB of memory. Its weights are available under Poolside's OpenMDW-1.1 license.\n\nThe challenge targets the MLX implementation used to run Laguna XS 2.1 on Apple Silicon. Eigen Labs said its internal agents had already produced a 36.8% speed improvement in the 24 hours before the public release on July 28th. The leaderboard shown by Naik four days later put the improvement at 138.7%, equivalent to about 2.39 times the original throughput under the challenge's benchmark.\n\nThat figure measures a specific inference implementation and workload. It does not indicate that Laguna's model weights became more capable, or that every Mac deployment will see the same gain. The result shows how quickly a shared, machine-checkable target can turn scattered optimization work into cumulative progress.\n\nPoolside co-founders Eiso Kant and Jason Warner have already built internal systems around automated experimentation. In its [Laguna S 2.1 release](https://poolside.ai/blog/introducing-laguna-s-2-1), Poolside described an automated loop in which its model optimized the company's own agent harness, cutting wall-clock time by 5.2% and memory allocations by about 70%. MLX.fast exposes a similar loop to outside engineers and their agents.\n\nEigen Labs now [describes its work](https://www.eigenlabs.org/) as building coordination technology for humans and AI agents. MLX.fast gives that strategy a commercial test: whether an open network can improve an open model's serving stack faster than the model developer or a hardware vendor working alone.\n\nThe early leaderboard establishes that participants can find substantial gains on a bounded benchmark. The harder test comes after the first wave of obvious optimizations, when improvements become smaller and require deeper changes to kernels, memory use and model execution. The platform's value will depend on whether it continues compounding after those easy gains are exhausted.", "url": "https://wpnews.pro/news/eigen-labs-agent-challenge-speeds-poolside-s-laguna-by-138-7-on-macs", "canonical_source": "https://runtimewire.com/article/eigen-labs-poolside-laguna-mlx-fast-agent-optimization", "published_at": "2026-08-02 16:01:26+00:00", "updated_at": "2026-08-02 16:57:42.000853+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-agents", "ai-research", "ai-tools", "generative-ai"], "entities": ["Eigen Labs", "Gajesh Naik", "Poolside", "Laguna XS 2.1", "MLX.fast", "Apple Silicon", "Google", "Sreeram Kannan"], "alternates": {"html": "https://wpnews.pro/news/eigen-labs-agent-challenge-speeds-poolside-s-laguna-by-138-7-on-macs", "markdown": "https://wpnews.pro/news/eigen-labs-agent-challenge-speeds-poolside-s-laguna-by-138-7-on-macs.md", "text": "https://wpnews.pro/news/eigen-labs-agent-challenge-speeds-poolside-s-laguna-by-138-7-on-macs.txt", "jsonld": "https://wpnews.pro/news/eigen-labs-agent-challenge-speeds-poolside-s-laguna-by-138-7-on-macs.jsonld"}}