{"slug": "local-open-weight-llms-in-coding-harnesses", "title": "Local Open-Weight LLMs in Coding Harnesses", "summary": "Local open-weight large language models (LLMs) with 30 billion parameters and a mixture-of-experts architecture achieve roughly 40 tokens per second on a Mac or DGX Spark, matching GPT 5.5 Pro subscription speeds and solving challenging coding problems. The choice of coding harness significantly impacts token usage, with Claude Code consuming twice as many tokens as Codex for the same tasks.", "body_md": "# Local Open-Weight LLMs in Coding Harnesses\n\nI have been taking different local open-weight LLMs for a test drive in different harnesses (Qwen-Code, Codex, Claude Code).\n\n30B Mixture-of-Experts models are kind of a nice sweet spot and can solve challenging problems. And they get roughly 40 tok/sec on a Mac or DGX Spark, which is similar to GPT 5.5 in a Pro subscription and totally usable for everyday work.\n\nMore interesting is also the harness choice! Claude Code seems to be using 2x as many tokens as Codex.\n\nGemma 4 E2B is here just for reference to show that the tasks can’t be trivially solved by smaller models.\n\nThe longer write-up is now available at [Using Local Coding Agents](https://magazine.sebastianraschka.com/p/using-local-coding-agents).\n\nSource: lightly edited website version of my [Substack note](https://substack.com/@rasbt/note/c-283141629).\n\n## Read Next\n\n[Inkling: A New Open-Weight 975B MoE with a Few Surprises Short note on Thinking Machines Lab's 975B Inkling open-weight model, its benchmark profile, sparse MoE design, short convolutions, embedding RMSNorm, and](/blog/2026/inkling-architecture-benchmark-notes.html)\n\n[200,000 Subscribers Short note celebrating Ahead of AI reaching 200,000 subscribers.](/blog/2026/ahead-of-ai-reached-200000-subscribers.html)\n\n[GPT 5.6 Has 72 Possible Configurations. What's A Good Default? Short note on how GPT 5.6 model and effort choices map onto training-time and inference-time scaling, producing 72 configurations.](/blog/2026/gpt-5-6-configurations.html)", "url": "https://wpnews.pro/news/local-open-weight-llms-in-coding-harnesses", "canonical_source": "https://sebastianraschka.com/blog/2026/local-open-weight-llms-coding-harnesses.html", "published_at": "2026-06-26 09:42:42+00:00", "updated_at": "2026-07-16 15:01:45.049785+00:00", "lang": "en", "topics": ["large-language-models", "developer-tools", "ai-tools"], "entities": ["Qwen-Code", "Codex", "Claude Code", "GPT 5.5", "Gemma 4 E2B", "Mac", "DGX Spark"], "alternates": {"html": "https://wpnews.pro/news/local-open-weight-llms-in-coding-harnesses", "markdown": "https://wpnews.pro/news/local-open-weight-llms-in-coding-harnesses.md", "text": "https://wpnews.pro/news/local-open-weight-llms-in-coding-harnesses.txt", "jsonld": "https://wpnews.pro/news/local-open-weight-llms-in-coding-harnesses.jsonld"}}