AI News

Trending story

Hands-on Test of Siltok Token Station's Local Inference Performance

1 reports1 story updatesUpdated Nov 13 14:10

Understand this story

AI overview

On November 13, 2026, China.com Technology published a hands-on report on the silicon-based token Siltok Token Station. The device's 16GB VRAM standard edition is priced at over 20,000 yuan and relies on the self-developed Siltok Speed inference engine to run large multimodal models locally; the report claims its GPU compute utilization (MFU) rises from 56.2% with conventional solutions to 95.5%. In the hands-on test, Qwen3.8-Flash-Next (approximately 176B parameters) achieved 96-100 tokens/s in text generation, about 2.1 times the 47 tokens/s of the comparison device; the full MiniMax H3 version generated a 768p 5-second video at only 0.023 yuan/second when depreciated over 3 years, 95% lower than the cloud API's 0.5 yuan/second. The device comes preinstalled with 9 major models including MiniMax H3 and LTX 2.5, and natively supports the Windows x86 ecosystem. There are currently no further developments.

AI-generated from public reports

Latest development · Report

Siltok Token Station hands-on test: a 16GB VRAM device priced at over 20,000 yuan delivers million-class AI workstation performance

Developments

1 confirmed developments

  1. Benchmarking Siltok Token Station's local inference performance

    Source:China.com Technology

Reports and discussion timeline

Browse reports and discussion by time. Their evidence levels remain distinct.

Nov 13

  1. China.com Technology

    Siltok Token Station hands-on test: a 16GB VRAM device priced at over 20,000 yuan delivers million-class AI workstation performance

    The Siltok Token Station 16GB VRAM standard edition is priced at over 20,000 yuan. Powered by the self-developed Siltok Speed inference engine, it runs large multimodal models locally, raising GPU compute utilization (MFU) from 56.2% in conventional solutions to 95.5%. In hands-on testing, Qwen3.8-Flash-Next (about 176B parameters) reached 96-100 tokens/s in text generation, roughly 2.1 times the 47 tokens/s of the comparison device; the full version of MiniMax H3 generating a 768p 5-second video costs only 0.023 yuan/second when depreciated over 3 years, 95% lower than the cloud API's 0.5 yuan/second. The device comes preinstalled with 9 major models including MiniMax H3 and LTX 2.5, and natively supports the Windows x86 ecosystem.

Related stories

  1. Related storiesQwen Releases Qwen4-Max Preview Version
  2. Related storiesThinking Machines Lab releases Inkling 2 Small
  3. Related storiesMistral releases Large 4 weights on Hugging Face
  4. Related storiesMeta releases open-weight model Muse Spark Open