Open-weight coding models get closer to closed systems

Frontier open-weight coding models are now competitive with closed systems on coding benchmarks, with DeepSeek V4-Pro, GLM-5.2 and Kimi K2.6 identified as the leading families. Top open coders clear 70-80% on SWE-bench Verified, while Qwen3-Coder stands out as the strongest downloadable option: the 480B-A35B flagship is described as open state-of-the-art, and the 80B-A3B variant keeps ~96% of that quality in a smaller workstation profile.

Hardware remains the main constraint. DeepSeek V4.1-Pro, GLM-5.2 and Kimi K2.6 sit in the 270-400 GB frontier tier and generally require 80 GB-class multi-GPU systems, big unified memory or API access. More practical self-hosted options include Qwen3-Coder 80B-A3B, Devstral-2 123B and MiniMax M3 in the 48-130 GB workstation tier, while Qwen 3.6 27B at Q4 is presented as the best local coder for a 24 GB card.

The recommended setup pairs a fast local model with a frontier API model. Qwen 3.6 27B handles chat, refactors and quick edits locally, Qwen3-Coder 8B supports fill-in-the-middle autocomplete, and tools such as Aider, Continue, Cline and OpenCode provide the agent loop needed to approach benchmark performance. GLM-5.2 or DeepSeek V4.1-Pro is reserved for the hard 5%, avoiding the cost of a 400 GB rig used at 5% duty.