Ask HN: Has anyone replaced Claude/GPT with a local model for daily coding?
TL;DR Highlight
Developers report running Qwen3.6 locally for much of their daily coding, with trade-offs.
Who Should Read
A developer weighing a local LLM coding agent to cut subscription costs and keep code private
Core Mechanics
- Only comments, no article body
- The post has no body, so the only source fact is the title question.
- Details are summarized in the community reaction section.
Community Reactions
- Qwen3.6 was the most common pick
- Many commenters run Qwen3.6 27B or 35B-A3B with llama.cpp and harnesses like OpenCode or Pi.
- One says Qwen 3.6 27b handles about 90% of their coding, with Codex for complex work and UI polish.
- Another replaced a $100/month Claude plan, saying it is not as good as Claude but not much worse.
- Frontier plans, local executes
- One commenter has Opus write plans that a local agent follows, then Opus validates them.
- Another uses Sonnet for design and a detailed plan, then feeds Qwen3.6 27b step by step.
- One runs overnight fuzz testing with the Chrome devtools MCP, a workflow that would be too costly on frontier models.
- Local models need precise instructions
- One says Qwen3.6 35b takes the easiest route when assumptions are open, loops often, and gets edit tool calls wrong.
- The same comment says Opus gives a 15x speedup while local Qwen gives 5x.
- Another says quality and speed drop past 100k context, so they set compaction at 75% of a 256k window.
- Skeptics also weighed in
- One tried every open model up to Qwen 480B and Kimi, and none came close to Claude.
- Another says setup time and cost are not worth it after researching monthly.
- One uses DeepSeek v4 Pro and Kimi K2.6 via hosted providers for 80%+ of sessions.
How to Apply
- Run a model that fits your GPU today
- With about 24GB VRAM, serve Qwen3.6 27B or 35B-A3B with llama.cpp.
- Connect OpenCode or Pi and give it one small bug fix.
- A commenter's setup guide is at https://github.com/pierotofy/LocalCodingLLM/.
- Split planning and execution
- Have a large model write the design and step-by-step plan once.
- Feed the plan to the local model one step at a time with precise prompts.
- Send only stuck tasks back to Claude Code or Codex.
- Lower your context compaction threshold
- If quality drops past 100k, set compaction around 75%.
- Keep tasks narrow and state assumptions explicitly.
Try It
# Paste into a local model (OpenCode/Pi etc.) or Claude — a simple version of the commenters' experience
You are a coding agent that only does narrowly scoped, clear tasks.
Rules:
1. If any assumption in the task is unclear, ask first. Do not silently take the easiest route.
2. Follow the architecture rules: {{project rules, e.g. keep CSS in separate files}}
3. If an edit tool call fails, retry the same edit once before re-reading files.
4. Do one step at a time, then summarize changed files and test results.
Execution plan: {{step-by-step plan from a larger model}}
Current step: {{step number and description}}Terminology
Related Papers
Cross-request KV cache exposure in the GLM-5.3 Workers AI model (155 responses)
A KV pool smaller than the context window leaked other customers' data into 155 long-prompt responses.
Microsoft Decision-1: 9B Qwen Model Launch Explained
Microsoft's 9B Qwen-based model returns only per-option probabilities for routing and classification.
Cerebras Deploys Trillion-Parameter AI Models
Cerebras fits Kimi K3 (3T) across 59 wafers in a pipeline, serving 1000-2000+ tokens per second
A practical guide to AI image generation, from diffusion to agents
Testing image APIs showed you pay per canvas, and camera-style prompts beat describing the view
DeepInfer: High-performance inference framework for DeepSeek models
DeepSeek open-sourced DeepInfer, its production inference engine, under Apache-2.0.
Mercury 3: Diffusion at scale
Scaling diffusion LLM 40x gives Mercury 3: 4–8x faster than speed-tier models, similar quality
DeepSeek banned from U.S. federal civilian agencies under new DHS order
DHS order bans DeepSeek from U.S. federal civilian agency IT systems
Timestral: Mistral's time-series foundation models
A 209M time-series model ranks first on three benchmarks and runs on CPU.