Free browser-based Jetson farm runs Cosmos3-Edge on 8GB via INT4 at 152 tok/s
No-Dragonfly6246 · reddit · 2026-10-23
Embedl offers a free browser-based device farm of real Jetson boards for testing Hugging Face models with runtime stats, supporting 1,000+ models.
Highlight: Cosmos3-Edge previously couldn't run on Orin Nano Super 8GB because the BF16 stack didn't fit — INT4 quantization now makes it work with peak VLM memory of just 3.3GB, while accuracy across five VLM benchmarks barely moved (65.56 → 65.20). Decode throughput: 152.5 tok/s on Orin Nano Super (optimized) and 207.8 tok/s on AGX Thor 128GB, versus 44.1/68.2 tok/s under BF16 vLLM. VLM is live; diffusion and policy components coming soon.
Related event: Embedl Opens Free Browser-Based Jetson Testing Platform(2 posts)→
More from Infra
- The Jensen effect: MicroDucks sells 4,000+ units in 24h after GTC showcase — andimarafioti · 2026-10-23
- Terry Tang co-founds TensorScale AI to rebuild inference infra for world models — TerryTangYuan · 2026-10-23
- Quantizing a 9B model: 70% of damage sits in the first 8 layers, and perplexity hides the tail — DinnerPrestigious953 · 2026-10-23
- HFT giants now fight for compute: Jane Street-backed Thinking Machines hits $40B valuation — dinabass · 2026-10-23
- Bloomberg: Tencent, Legend Capital and CPE circle Huawei Ascend chips, same names near DeepSeek — YvesMulkers · 2026-10-23
- How Many AI Agents Can a $6/Month 1GB VM Run? They Measured 200 at Once — swarmagent · 2026-10-23