A practical warning for using Hermes on local coding work
An M1 128GB setup was used to test whether could replace advanced paid models for many coding tasks, but the results fell short. The first goal was to replace about 90% of advanced model work with , but that proved unrealistic. A second approach used an advanced model for planning and a local model for execution, but that also did not work reliably.
Several coding tools were tried, including opencode, openhands, picode, and Hermes. The breaking point was a board task where Qwen 27B spent millions of tokens trying to fix a table div and still failed. Even with Q8 and , many requests reached about 100,000 to 120,000 tokens, while the model still lost context and made unwanted changes.
Local models worked better for one-shot jobs like summarization and than for long coding-.
Key points
- An M1 128GB machine still did not make feel like a broad replacement for advanced paid models.
- Hermes, opencode, openhands, and picode were all tested for coding-agent style work.
- A board run with Qwen 27B failed to complete a simple table div fix after using millions of tokens.
- Q8 and did not prevent context loss on 100,000 to 120,000 token requests.
- One-shot tasks like summarization and worked better than long coding workflows.