Open-weight shortlist from Kimi 2.8T down to Qwen 27B: practical beats smartest
drdanielbender · x · 2026-11-13
X user drdanielbender published an open-weight model shortlist thread, arguing the most interesting AI model isn't always the smartest one — it's the one that runs compliant in your region or on a GPU you already own.
- Includes an Artificial Analysis Intelligence Index chart of open vs proprietary models: the closed frontier (Claude/Gemini/GPT, 65-54 points) leads, while Kimi K3 Pro scores 46, Qwen3 235B 34, DeepSeek V3.1 Flash 32
- His shortlist covers 4 models from Kimi's 2.8T parameters down to Qwen's 27B, balancing capability and local-deployment cost
- He highlights Qwen3.8-27B (fits 24GB GPUs at Q4) and notes Qwen4's open-weight release is imminent
Related event: Four Open-Weight Models Worth Running Locally, from 27B to 2.8T(4 posts)→
More from Infra
- Back-of-envelope: serving 1B AI Mode queries on Gemini costs only ~$4M — gaganghotra_ · 2026-11-13
- After Jensen Huang's resignation, Korea convenes emergency meeting on its 260,000-GPU NVIDIA plans — JungWooHa2 · 2026-11-13
- Same Cloud Machine Size, Nearly 2x Speed Gap: A Speech Recognition CPU Lottery Postmortem — shauntrennery · 2026-11-13
- GPU stagnation: NVIDIA's Blackwell called last GeForce, IPC gains at zero — soham_btw · 2026-11-13
- NVIDIA's six main suppliers grew 64% YoY last quarter, fastest in a year — tengyanAI · 2026-11-13
- Nvidia supply chain: 2027 orders unchanged after Huang exit, some run to 2028 — firstadopter · 2026-11-13