1-Bit Compressed 295B Model Outperforms Cloud API Locally
iamaliveix · x · 2026-07-15
Atomic Chat applied extreme 1-bit compression to the 295B parameter Hy3, generating a 92GB file that runs locally on 4× RTX 5090 / 128GB VRAM, achieving speeds 2.2x faster than cloud APIs.
In testing, both the local and cloud API versions completed the same one-shot prompt tasks, generating games like Flappy Bird, Arkanoid, and Snake without noticeable quality drops or crashes. Their output scales were similar: the local version output 76.9K tokens / 15.5 minutes, while the cloud API output 75.1K tokens / 34.3 minutes.
The author emphasizes that Hy3 is an Apache 2.0 model open-sourced by Tencent, available for commercial use, and is a large 295B MoE (21B active) model.
Related event: Atomic Chat Launches 1-Bit Hy3 Offline Chat App(3 posts)→
More from Infra
- SF Compute founder: buying compute is 'an absolutely awful experience' right now — IgorCarron · 2026-09-11
- SmolVM open-sources persistent computer infrastructure for agents that outlive chat sessions — aniketmaurya · 2026-09-11
- PyTorch Day Korea 2026 launches first offline conf, CFP closes Sept 13 — PyTorch · 2026-09-11
- Local LLM server dilemma: 4x CMP-170HX (price up 53% in 20 days) vs Mac Studio M5 Ultra — rumboll · 2026-09-11
- llama.cpp lands Flash Attention tuning for RDNA4, big prefill gains on AMD — pmttyji · 2026-09-11
- Your p99 latency benchmark may be lying: a deep dive into coordinated omission — Franc0Fernand0 · 2026-09-11