Democratizing Intelligence: Building an Open Source AI Lab on a Decentralized Network

Should intelligence be owned by the broligarchy?

Intelligence has quickly become one of the most critical resources on earth, and unfortunately it has already turned into a race for unimaginable wealth and concentrated power. A small handful of companies own the frontier models, the data, the hardware, and the IP to produce these models. Those same companies are now pushing for regulatory capture and anti-trust wavers to further consolidate control. The durable solution isn't just another open weight model with secret proprietary code owned by a company seeking a high valuation. The durable solution is an open, unstoppable intelligence platform that is fully open source, and powered by truly decentralized compute that anyone can contribute to. A platform with permissionless participation, creating new models explicitly designed to be efficient locally, comfortably running on commodity hardware.

Democratizing intelligence is the mission. It outranks everything else we do and the urgency has never been higher. With the calls for banning development on AI or pushing for regulatory capture in the news almost daily, this open AI model producing network MUST be built. Chutes sees this absolute need and is accelerating efforts to make it a reality. The first step in doing so is a new paradigm of pretrained base models with an architecture designed to explicitly target high speed inference on consumer grade hardware. These models aim to be at the absolute frontier, pushing limits of efficiency and compactness. We have the algorithms, we have the data, we've achieved incredible results. Now we must accelerate and build bigger and better models.

Chutes Subnet 64 started as serverless compute, and we used inference to showcase the capabilities of the network which then grew into a real business. That business continues, and now is more important than ever because it powers our efforts to build highly capable new models. It also serves as a powerful distribution channel for accessing inference from the models we will produce.

1. Open source will win

Most pretraining data is basically the entire written record of humanity. Ownership and access to the models should reflect that. Instead, we get companies that violate copyright laws, train on user data, and do not contribute to open source in a meaningful way despite their non-profit foundational mandate. Consider their pitch for instance that there are many TRILLIONS of dollars in total addressable market. Thats the entire GDP of the United States. These companies are racing to replace all human labor and further consolidate all information into moated services with exceptionally few beneficiaries. Simultaneously they seek to pull up the ladder behind them restricting open source and forcing regulation etc.

Look at who is asking for that trust:

• Labs that could not, for two days, rule out that a mathematician's private prompts had trained their model (it should never have been a possibility).
• Labs that settled copyright claims for billions of dollars, because the fee was worth the theft and irrelevant to the valuation.
• Labs settling for billions of dollars for often harmful ads targeting vulnerable populations.
• Labs facing wrongful death suits and harm over what their "aligned" models said to people in crisis.
• Labs whose customer traffic was quietly routed through thousands of shell/fraudulent accounts to perform industrial scale distillation/data collection.
• Labs that did not or could not build proper sandboxing environments resulting in what could be considered felony level cybercrime activity.
• Labs and companies that are continuously consolidating and folding up tools and platforms developers and researchers rely on to further open source efforts.

Beyond this, the risk of "poisoning" the models from any lab, open weight or not, remains. "sleeper agents" can be trained into the data with only a few hundred poisoned documents, planting a trigger in a model of any size. A model can carry trained-in (opinionated) preferences, product advertisements, a missing decade of history, etc., and no benchmark will tell you. Even accepting this risk, export controls can delist the weights you depend on overnight, in either direction. A defense bill names a Chinese lab, and a Chinese ministry consults on blocking foreign downloads of Chinese weights.

Do we have a solution to all of these problems? No; no model will be perfect! But this is precisely why the pipeline has to be fully open. All weights, data, tokenizer, training code, sync algorithms, evaluation, and the receipts that tie them together so the entire planet of AI users and researches can examine, improve, and iterate. Not "open weights" with a mystery corpus, not "open" with a license that expires when you succeed. Truly open so that you know every single ingredient that went into baking the model, fully auditable and improvable. Open so the result is efficient and cheap enough to run on hardware you own instead of a rack of B300s rented from a hyperscaler.

There has never been a more critical moment, a necessary inflection point in pushing open source AI. The frontier is expensive; but with efficient systems we can make it so catching up is not. This needs to be built now, not tomorrow when the regulatory tides have changed rendering it impossible.

2. Corporations are Already Moving to Sovereign AI

Large corporations are not waiting, they are moving their intelligence in-house. Today, bringing intelligence in-house means access to massive data centers and the ability to spend hundreds of millions of dollars. Most businesses want private, local AI. Almost none can afford to build it.

A few recent examples:

Latham & Watkins, the second-largest US law firm, bought its own Nvidia GPU servers and is fine-tuning open-weight models. They plan to spend $500M for the hardware and staff for on-prem GPU clusters.

Cursor, Intercom and Cognition now run their core products on models they post-trained on open weights. Expect them to train their own models in the very near future.

Eli Lilly built the largest AI factory in the Pharma industry and acquired more than 1,000 Blackwell GPUs for training models on their own data.

Chinese companies whose core business has nothing to do with AI are training their own foundation models: Meituan (food delivery), Rednote (social commerce), Xiaomi (phones and cars), Ant Group (payments) and Kuaishou (short video). Meituan's LongCat-2.0 is a 1.6 trillion parameter model trained on its own cluster of 50,000 Chinese chips.

The corporations have already decided they can't trust their intelligence to someone else. Why should you?

3. Decentralized training hit a wall

The honest state of the field in September 2026:

• Several "decentralized" models turned out, after analysis, to be forged updates distilled from existing open weights. It's unfortunate, but a difficult problem to solve, because the best weight updates to publish as a miner are the ones that lower loss, and the easiest way to guarantee that is to forge a synthetic update from pre-existing models. Loss is not monotone under hard data, so the honest update looks worse than the forged one.
• Validation schemes are incredibly difficult, and it's virtually impossible in most designs to stop one miner from poisoning the shared model.
• The infrastructure is often not decentralized. Cloudflare R2 or similar object store required by miners and validators uses hardware only found in datacenters/hyperscalers, etc. A small number of off switches.
• In some cases, the networks have stalled or been put on indefinite hold. One team moved post-training onto a co-located cluster, some have totally rugged the network that built them and walked away.
• MoEs are another class of challenge, in that each island may learn different expert specialization and therefore merging results causes redundancy and lack of expert differentiation.
• Decentralized training just for the sake of decentralized training is pointless. What we need is a durable, unstoppable system for building models (and not just a model, a model with a purpose).

4. We accept the challenge.

Parallax is how we train models across the internet without a data center. We have completed multiple full runs on a live fleet of consumer gaming GPUs. The most recent published run was an 8B parameter model finished 10 October. It processed 901 billion tokens in under six days and used less than $7k in compute. This model also shipped at just 2GB in total size. Normally models in its class need 17GB or more and require specialized hardware.

Training costs just $11.27 per billion tokens on our current fleet. That rate holds at whatever token budget you want. Our next model is significantly larger and already in progress.

It’s a very complex set of algorithms combined into one extremely powerful framework. The training details and specifics of how the method works can be found here Parallax · live training register The final state as we intend it will be signed membership with unbounded joins at any step. A hash-chained ledger of every update, data provenance hashes on every streamed shard, signed package updates with slow rolling upgrades of the whole fleet at runtime, etc. And we can prove the recipe and code with attested workloads using our existing TEE attestation mechanisms. Poisoning and forging updates will be a thing of the past.

Ultimately what we care most about is ubiquitous tokens and intelligence. That means producing highly efficient and compact models that don’t require data centers. Models designed for local inference. These models aren’t currently being produced, and we don’t have a massive collocated cluster to train them traditionally, so decentralized training is just a road we had to build to get there. Ternary experts with structured sparsity for ultra-fast experts, even the potential to port to CPU lookup-table kernels, which could potentially fit a trillion parameters in ~200GB. This is what "accessible" means to us. Fine tune models on a workstation, run it on edge/local hardware, no outages because someone accidentally hit a fiber line. No risk of data leakage. No IP theft. No $550k capex to get "local" inference on an 8x b300 box.

5. How we get there

The network points at the mission

Over time the network's incentives will be gradually shifting toward focusing more heavily on training, SFT, RL, data pipelines, etc to ensure there is a clear link between the network, our community and open source development. Direct pipeline into open, extraordinarily efficient democratized intelligence.

This is an additive move. Existing miners will be unaffected. We are simply opening a new lane for consumer hardware that could never have served production inference. Today only those with datacenter grade hardware can provide inference on Chutes. Instead we want a network where the bar to entry is affordable to every day people.

The Chutes Network’s inference infrastructure is a core part of how we get to truly open intelligence. Decentralized inference across the planet but reliant on data centers was a good first step. Now it’s time for the next step where we decentralize the very kitchen that cooks the models. Break the dependency on datacenters and put the power in the hands of everyone and anyone willing to contribute.

Imagine a global inference & training network where consumer grade hardware like a DGX Spark or Apple silicon in peoples' homes support the network (alongside GPUs within datacenters).

We will publish models, checkpoints, datasets with lineage, training and sync code, evaluation harnesses and benchmark results, ablations and postmortems, etc.

The platform carries it

Customers, inference access and custom chutes keep working. The confidential (TEE) end-to-end encrypted network stays, with a catalogue of models that can be run sustainably with no wasted idle compute. Private chutes, confidential inference, etc. remain products in their own right, because they are critically important to data pipelines needed to produce the models, and we'll host these models on the inference platform we've already built (in addition to being published on HF).

We're sourcing additional non-TEE inference to expand our offerings for our existing customers, paid from revenue not emissions.

The time for change is now. The threats are apparent but the goal has never been clearer. We are on our way and we hope that you will join us in this grand objective to bring intelligence to all, unrestricted and open for all mankind.

6. Q&A

Is Chutes still an inference platform? Yes. We will continue to serve inference, including TEE models with attestable open source end to end encryption.

In addition, we will soon be offering additional compute for inference that does not utilize network emission.

 What happens to miners? Nothing changes for miners today. Existing workloads, hardware and earnings continue as they are. Over time we open a new lane for consumer hardware (DGX Spark, Apple silicon) to contribute to training, weighted separately from GPU compute. Miners who want to participate in both can.

 Why not just fund a closed lab and open the weights at the end? Because again, a weight file with no provenance and magic formula/data is a black box with a license. Immense amounts of trust are required even for open weight models.

Why now? Open models are closer to the frontier than they've ever been, consumer hardware has never been this capable, and the open framework must exist before AI cabals decide it cannot. Frontier AI labs are trying to halt progress and put up barriers. Must accelerate.