Bittensor Ecosystem Highlights (July 13–19, 2026): Parallax Nears Datacenter Baseline, Leadpoet Lands on OKX, and RedTeam Bot Detection Hits 99.5%

Decentralized training closed the gap on centralized baselines this week, one subnet made history as the first Bittensor product on OKX’s agent marketplace, and confidential compute expanded across live GPU inventory alongside standout security and detection benchmarks.

Here’s the rundown.

Subnet updates

Chutes (SN64) trained a 20B model on cheap, mismatched GPUs to within about 2% of a datacenter baseline using Parallax (source).

Leadpoet (SN71) became the first Bittensor subnet announced for the OKX AI Agent Marketplace (source).

RedTeam (SN61) reported its bot detection rate rose from 74.3% to 99.5% over 11 months of live attacks (source).

Lium (SN51) reported record-high revenue and usage, with both usage and GPU supply more than doubling over the past month. The team also launched confidential computing on select GPUs, using TEEs to encrypt data and restrict access from GPU providers (source, source).

GM (SN28) launched on mainnet, offering a TEE-based API compatible with existing OpenAI, Anthropic, and Gemini code (source).

404 (SN17) generated 21.5M+ high-fidelity assets, making it the world’s largest open-source 3D repository by asset count (source).

BitMind (SN34) reported that its BMF deepfake detector matched or exceeded the strongest published results across seven public benchmarks without benchmark-specific tuning (source).

Beam (SN105) transferred 1.12 TB across its decentralized bandwidth network in 3 minutes and 40 seconds (source).

Swarm (SN124) launched SOTA Pilot v5, expanding its miner competition into multi-drone coordination and high-speed autonomous interception (source).

Score (SN44) is building the interface and API for users to access and subscribe to the decentralized VLM miners it plans to train on SN44 (source).

Takeaway

Parallax training a 20B model within 2% of a datacenter baseline on cheap, mismatched GPUs is the compounding technical win of the week, and the kind of result that repositions what decentralized training can actually deliver at scale.

BitMind matching top deepfake detection benchmarks without any benchmark-specific tuning is another reminder that specialized decentralized systems can hold their own against centralized generalists on the right problems. Additionally, Lium doubling both revenue and supply in a month while shipping confidential computing on the same infrastructure is the pattern that keeps repeating: subnets running like businesses with the cash flow to back it up.

On to next week.