Table of Contents
Quasar, the Bittensor SN24 project focused on long-context AI, has outlined why it is prioritizing model architecture before expanding its product stack.
In an article published on X, the team argued that memory is one of the most important constraints facing current AI systems. Here, memory refers to a model’s ability to retain and reason across very large working contexts without losing coherence or allowing important information to fall away.
The argument builds on Quasar's recent push around Quasar-Preview, a Mixture-of-Experts (MoE) checkpoint released by SILX AI for SN24. We previously covered Quasar's planned 10-trillion-token decentralized training run, which the team has positioned as a major test of whether decentralized infrastructure can support model improvement at larger scale.
Antonio Verrico
Quasar's Core Argument: AI Still Forgets Too Much
Quasar’s central claim is that many AI agents remain limited because their underlying models struggle to preserve useful context over long periods.
The industry often works around that problem at the application layer through retrieval systems, vector databases, summaries, checkpoints, and other forms of external memory. These tools can make AI applications more capable, but Quasar argues that they do not fully solve the underlying architectural limitation.
When a model repeatedly compresses, summarizes, or discards parts of its working state, information loss can accumulate. Over longer tasks, that can weaken reasoning, reduce consistency, and make agent behavior less reliable.
This matters most in applications that require sustained attention across large amounts of information. Software engineering agents may need to track files, dependencies, prior changes, and unresolved issues across an entire codebase. Research systems may need to compare papers, retain evidence, and revisit earlier assumptions. Enterprise agents may need to reason across policies, operational history, and interconnected workflows.
In each case, a large advertised context window is only valuable if the model can continue using that context effectively.
Quasar’s objective therefore goes beyond increasing token limits. The project is focused on active context quality: how well a model can retain, retrieve, and reason with information as its working context grows.
Why the Model Comes Before the Product
Quasar also used the article to explain why it has emphasized model development ahead of a broader product rollout.
The project already has a product direction in Quasar Copilot, an agent intended to operate across a user’s wider working context rather than inside isolated sessions. But the team argues that building the application around a conventional model would carry forward the same memory limitations it is trying to address.
A polished interface, more integrations, or stronger orchestration cannot fully compensate for a model that loses coherence as tasks become longer and more complex.
That is why Quasar describes the model itself as the project’s current product. Its strategy is to improve the foundation first, then build the application layer around a model designed specifically for long-horizon work.
The Architecture Behind Quasar-Preview
Quasar-Preview uses a 20-billion-parameter Mixture-of-Experts architecture with 2 billion active parameters.
In a Mixture-of-Experts model, only a subset of the network is activated for each token or request. This can allow a model to maintain greater total capacity without using every parameter for every operation.
Quasar says the design has produced strong long-context results in internal benchmarking. The team claims that the model retains roughly 90% of its performance at 100,000 tokens and approximately 84% at 1 million tokens.
Those figures have not yet been independently validated, but they illustrate the type of capability Quasar is trying to optimize: not simply accepting more tokens, but preserving useful performance as context expands.
A Decentralized Training Test for SN24
Quasar’s roadmap now points toward a much larger decentralized training effort.
The team says its 2026 goal is to complete an initial training phase using 5 trillion tokens, followed by another 5-trillion-token phase to reach 10 trillion tokens in total.
Quasar has also outlined plans to scale model capacity from its current roughly 20-billion-parameter range toward 40 billion parameters. Its longer-term ambition is to reach 100-billion-parameter scale while continuing to improve training quality and post-training performance.
That roadmap positions SN24 as more than a venue for a single model release. Quasar is attempting to create a repeated training and improvement system in which miners, evaluators, and researchers contribute to a shared model direction over time.
For Bittensor, the significance lies in the coordination model. Centralized AI labs rely on concentrated capital, compute, and closed infrastructure. Bittensor subnets instead use token incentives, miner competition, and open participation to organize specialized AI production.
Quasar is applying that model to one of the harder areas of AI development: training a frontier-scale model with durable long-context performance.
What Quasar Could Prove for Bittensor
Quasar’s importance to Bittensor is not simply that it is building another long-context model. The larger test is whether an incentivized subnet can coordinate repeated training, evaluation, and improvement around a shared frontier-model objective.
If SN24 produces models that developers and researchers choose to use outside the Bittensor ecosystem, it would strengthen the case that subnets can contribute to core model development rather than primarily supplying inference, data, or compute.
That remains unproven. Quasar’s benchmark claims need external validation, and its training roadmap must still translate into measurable model improvements and real adoption.
But the project is targeting a genuine AI bottleneck. More capable agents will require models that can preserve context, maintain reasoning quality, and operate reliably across longer and more complicated tasks.
Quasar’s bet is that those capabilities must be built into the model before they can be delivered convincingly through a product. If SN24 can demonstrate that decentralized incentives can help produce them, the result would matter well beyond a single subnet.
Disclaimer: This article is for informational purposes only and does not constitute financial, investment, or trading advice. The information provided should not be interpreted as an endorsement of any digital asset, security, or investment strategy. Readers should conduct their own research and consult with a licensed financial professional before making any investment decisions. The publisher and its contributors are not responsible for any losses that may arise from reliance on the information presented.
