Physix Frontier badgePhysix Frontier
PublishChinese | EN

Physix Frontier · News Briefing Card (Hacker News · Nov 5, 2026)

Microsoft ships local coding model: 137.9B MoE on RTX Spark

KEY FACTS

  • Microsoft released MAI-Code-1.1-Flash, a 137.9B-parameter MoE coding model.
  • Model weights were compressed from a 276GB 16-bit version to a 53GB mixed-precision build.
  • On RTX Spark, median decode speed is 51.12 tokens/s at a 120K context.
  • The model is live on Microsoft Foundry and offered as a local model in GitHub Copilot.
  • A smaller 46.1GiB variant averaging 2.87 bpw is also available, with precision assigned to expert layers by role.

KEY DATA

137.9BParameters
276GB16-bit weight size
53.0GiBRelease weight size
51.12 tokens/sDecode speed

PHYSIX OBSERVATION

Running a hundred-billion-class coding model locally is approaching usable for the first time: 53GB of weights plus 128GB of unified memory frees developers from per-token billing. But 51 tokens/s is only a median, and long-context and real-world completion experience remain to be verified. The competitive focus in low-bit quantization is shifting from parameter scale to precision allocation and VRAM scheduling, and this will be the new battleground for PC-side AI.

Source: Hacker News report