Bittensor This Week: Chutes Closes the Decentralized Training Gap to 0.6%, RedTeam Lands Its Largest Round

A week of quiet but meaningful milestones; a serious technical result on decentralized training, external validation from NVIDIA, and fresh capital into a security subnet.

Here’s the rundown.

Subnet updates

Chutes (SN64) achieved fully non-blocking decentralized training on a recurrent model with Parallax, staying within a 0.6% quality gap versus centralized training (source) — arguably the strongest single result yet in the Parallax arc.

Vidaio (SN85) launched Sentinel, a surveillance-focused product built on its video compression engine for security footage and CCTV workflows (source). Vidaio’s compression models are also being integrated into Manako AI so customers can cut bandwidth costs directly inside edge vision workflows (source).

Synth (SN50) published its first research paper using SN50 as a live benchmark for time-series forecasting in finance, featuring a Chronos-2 model fine-tuned on Synth (source).

Trishool (SN23) published the full HaloGuard 1.0 paper, covering its constitutional data pipeline, benchmark results, multilingual coverage, and red-teaming loop — a follow-up to last week’s SOTA announcement (source).

Score (SN44) — Manako joined the NVIDIA Inception Program (source).

Metanova (SN68) shared its H1 2026 update: NOVA is expanding into nanobodies, cleaner subnet incentives, agent tooling, robotic synthesis with OnePot, with first wet-lab results planned for H2 (source).

RedTeam (SN61) announced its largest funding round to date and a partnership with Unsupervised Capital (source).

ChronoLLM (SN38) — Crunch introduced ChronoLLM, a point-in-time model project where SN38 trains historical checkpoints, scaling toward 72B parameters (source).

Ecosystem moves

TAO.com rebuilt its Chrome wallet extension from the ground up for subnet trading (source).

Takeaway

Two threads run through the week.

The first is technical: Chutes closing the decentralized-versus-centralized training gap to 0.6% on a recurrent model is the kind of number that flips the conversation from “can it work?” to “is there any reason not to use it?” — especially paired with published research from Synth and Trishool that starts to give subnets an academic paper trail.

The second is external validation: Manako into NVIDIA Inception, RedTeam raising its largest round to date with Unsupervised Capital, Metanova moving toward wet-lab results. These are the kinds of signals that come from outside the ecosystem — and increasingly, they’re what the ecosystem is being measured on.

On to next week.