Following pricing changes from major cloud providers, Qwen 3.6-27B and DeepSeek V4 Flash have emerged as the top choices for local coding models across various VRAM tiers.

  • Qwen 3.6-27B achieves SWE-bench Verified scores comparable to Claude Sonnet 4.6 while running on consumer hardware.
  • DeepSeek V4 Flash offers low-latency tool-calling capabilities at Haiku-tier API rates.
  • Gemma 4 26B-A4B MoE provides a competitive alternative for local inference with efficient active parameter usage.
  • Speculative decoding methods like DFlash and MTP are now available in production runtimes to improve speed.

The Qwen 3.6 family remains the open local flagship, offering strong agentic coding performance for users seeking alternatives to cloud-based solutions.