5x for Free：本地编程栈
5x for Free : The Local Coding Stack
Hacker News 讨论揭示：Qwen 3.6 35B-A3B 模型提及率 33% 领先，27B 变体以 20% 紧随其后，DeepSeek Pro 与 Gemma4 31B 位列前四。Agent 工具中 Pi (49%) 与 OpenCode (45%) 占主导。用户对比称，Claude Opus 可带来 15 倍加速，而本地离线 Qwen 提供 5 倍加速，且完全免费、保护隐私。SWE-bench Verified 基准测试显示，Qwen 3.6 27B 得分 77.2%，35B-A3B 得分 73.4%，接近 Claude Sonnet 4.6 的 79.6%。MoE 架构使大模型在消费级硬件上高效运行。
本地模型在编码上正逼近云端前沿，Qwen 35B-A3B 已成社区标配，免费且完全离线让这场替代变得真实，选型逻辑可能从此改变。
In short : A 500+ comment HN discussion reveals the current state of local AI coding - Qwen 3.6 35B dominates model mentions while Pi leads agents.
Today, a Hacker News thread asked a simple question : “Has anyone replaced Claude/GPT with a local model for daily coding?”1 500+ comments later, a clear picture emerged of the local coding stack.
Qwen 3.6 35B-A3B dominates model mentions at 33%, followed by the 27B variant at 20%. DeepSeek Pro & Gemma4 31B round out the top four. The common thread : mixture-of-experts architectures that run fast on consumer hardware.2
On the agent side, Pi leads at 49% with OpenCode close behind at 45%. Both are lightweight harnesses designed for local inference.
The thread surfaced a fascinating tradeoff. One commenter captured it perfectly :
Comparing agentic Qwen3.6 35b to Claude Opus is like a junior with knowledge across the board, that you really need to guide, versus a senior that thinks with you on architecture. If Opus gives a 15x speedup, local and fully offline Qwen gives a 5x speedup.
But for many, the tradeoff is worth it. Privacy, zero cost, & complete offline capability matter.
Given that it’s completely free, is still mind-boggling to me.
The local coding stack is maturing fast. Qwen 3.6 35B-A3B has become the de facto standard & Pi the leading harness.
The benchmark data backs up the sentiment. Qwen3.6 27B scores 77.2% & the MoE variant, Qwen3.6 35B-A3B, hits 73.4%. These two local models are within spitting distance of Claude Sonnet 4.6 (79.6%).3
This is the minimill pattern playing out in real time. It’s not just for CRM updates & web research. The current generation of local models is good enough for reasonable coding tasks.
-
MoE models are large models that only activate a small fraction of their total parameters. Qwen 3.6 35B-A3B has 35 billion total parameters but only 3 billion active at inference time, while the 27B variant runs all 27 billion parameters each time. ↩︎
-
SWE-bench Verified scores from llm-stats.com & morphllm.com, June 2026. ↩︎
来源：Tomer Tunguz 博客（VC 分析） · tomtunguz.com