Table of Contents
Chutes plans to build a fully open-source AI lab on Bittensor Subnet 64, expanding the network beyond serverless inference into training models that run efficiently on consumer hardware.
The team will gradually direct more network incentives toward model training, supervised fine-tuning, reinforcement learning, and data pipelines. Its existing inference business will continue, helping fund development and distribute the resulting models through a platform customers already use.
The open-source lab plan builds on Parallax, Chutes' framework for training across geographically distributed computers. Chutes says its latest published 8-billion-parameter run processed roughly 901 billion tokens in less than six days on consumer gaming GPUs, using less than $7,000 in compute.
Chutes is targeting pretrained base models whose architecture prioritizes fast local inference, the process of running a trained model to generate outputs. The aim is to make models practical to operate on hardware people own, rather than requiring a rented data-center server.
That objective extends beyond releasing downloadable weights. Chutes plans to publish checkpoints, datasets with lineage, training and synchronization code, evaluation tools, benchmark results, ablation studies, and postmortems. Dataset lineage records where training material came from; ablations test which design choices contributed to a result. Together, those materials would let developers inspect and reproduce more of the model-building process.
In its announcement on X, Chutes argued that control over models, data, and hardware has become concentrated among a small number of companies. Its proposed alternative combines an open development pipeline with a broader pool of compute contributors.
The direction builds on Chutes' earlier efficiency roadmap, which emphasized reducing model memory requirements and improving useful output per GPU alongside a sustainable inference business.
What Parallax Has Trained
The Parallax training register records 901.3 billion tokens processed over five days, 17 hours, and 36 minutes for the lambda run, which ended October 10. The fleet launched with 208 RTX 5090 GPUs across 26 hosts in 11 countries, coordinated over the public internet.
The published lambda model card puts the model at approximately 7.77 billion total parameters, with about 1.20 billion active for each token. It is a mixture-of-experts model: a router selects a subset of specialized neural-network components for each token instead of activating the entire parameter bank.
Lambda combines recurrent processing with sliding-window and global attention. Its routed experts use ternary weights, represented by minus one, zero, or plus one, alongside scaling factors. Structured sparsity further restricts which weights can be nonzero. Those choices reduce the storage and arithmetic required for the expert layers.
Chutes summarizes the model as roughly 2GB; the model card lists its final checkpoint export at 2.43GB. These are stored-model sizes, rather than measurements of total runtime memory. The released checkpoint is a base model that continues text, not an instruction-tuned chatbot, and Chutes labels it a research artifact rather than a production release.
Parallax separates synchronization of the shared model backbone from updates to its experts. GPUs train expert adaptations locally, while expert owners incorporate those updates into full-precision master weights and publish new ternary versions. The backbone uses decoupled DiLoCo, allowing local training work between synchronization rounds instead of requiring every machine to exchange gradients after every step.
The architecture has evolved since the initial Parallax design, which distributed expert ownership and used lightweight approximations for remote experts. A subsequent recurrent-model experiment reported results within 0.6% of a centralized baseline at matched training steps while using non-blocking synchronization.
For the next generation, Chutes is exploring ternary experts with structured sparsity and possible CPU lookup-table kernels. It estimates that such an approach could eventually fit a trillion parameters into roughly 200GB. That is a proposed storage direction, not the size or demonstrated capability of lambda. A significantly larger model is already being trained, the team says.
Opening Participation and Verifying Work
The planned training lane would allow hardware such as NVIDIA DGX Spark systems and Apple silicon computers to contribute alongside data-center GPUs. Chutes currently requires data-center-grade hardware for production inference, so training would create a separate route into the network.
Training contributions will be weighted separately from GPU compute, and miners will be able to participate in both. Chutes says existing workloads, hardware requirements, and earnings remain unchanged today as it develops the new lane.
The incentive shift covers several stages of model development. Pretraining builds a base model from large datasets; supervised fine-tuning teaches behavior through examples, while reinforcement learning adjusts behavior using reward signals. Data-pipeline work supplies and processes the material those stages consume.
Chutes also plans signed membership, a hash-chained record of updates, provenance hashes for streamed data shards, and signed software packages rolled out gradually across the fleet. Its existing trusted execution environment, or TEE, attestation mechanisms would help verify the code and training recipe running inside protected workloads.
