Run a 125B MoE coding-agent model on two 16 GB gaming GPUs
capy-flash runs Qwen3.8-Flash-Next (125B MoE) in NVFP4 on two RTX 5070 Ti cards. In the author's tests it finished 20 of 20 multi-step coding tasks with zero malformed tool calls and serves OpenAI- and Anthropic-compatible endpoints.
Source
Feyd on Hacker News · 2026-10-16 · Project link
Written in our own words. Credit to the original author. Link checked 2026-10-16.
What they did
capy-flash runs Qwen3.8-Flash-Next (125B MoE) in NVFP4 on two RTX 5070 Ti cards. In the author's tests it finished 20 of 20 multi-step coding tasks with zero malformed tool calls and serves OpenAI- and Anthropic-compatible endpoints.
Result (as claimed)
About 100 tokens/sec generation in the author's benchmark; 7 points on Show HN.
Tools they used
Cost: Pricing depends on the tools you pick; check each official pricing page.
Set it up
Pick a contained repo or task
Start on a branch or side project where mistakes are cheap.
Install and configure
Set up Qwen. Read what it can run on your machine and limit permissions where possible.
Run it on a real task
Give it a well-scoped issue and let it work. Review the diff like any pull request.
Add guardrails
Keep tests and CI in the loop, and never let an agent push to main or touch production secrets unattended.
Copy the full setup prompt
Paste into your agent and it will walk you through the setup, asking before anything that sends, posts or spends.
You are helping me set up this AI use case end to end in Claude Code. Goal: Run a 125B MoE coding-agent model on two 16 GB gaming GPUs What the original author did: capy-flash runs Qwen3.8-Flash-Next (125B MoE) in NVFP4 on two RTX 5070 Ti cards. In the author's tests it finished 20 of 20 multi-step coding tasks with zero malformed tool calls and serves OpenAI- and Anthropic-compatible endpoints. Reported result (as claimed by the author): About 100 tokens/sec generation in the author's benchmark; 7 points on Show HN. Source: https://news.ycombinator.com/item?id=50118153 Tools they used: Qwen Category: Dev & ops Please: 1. Ask me 3 short questions about my situation (accounts I have, budget, what I want the output to look like). 2. Propose a step-by-step plan using these tools or cheaper alternatives, with the time and cost of each step. Use official pricing pages; say "check pricing" if unsure. 3. Then do the setup with me one step at a time, waiting for my OK before anything that sends messages, posts publicly, spends money or deletes data. 4. Finish with a checklist to run it again next time.
You are helping me set up this AI use case end to end in Codex. Goal: Run a 125B MoE coding-agent model on two 16 GB gaming GPUs What the original author did: capy-flash runs Qwen3.8-Flash-Next (125B MoE) in NVFP4 on two RTX 5070 Ti cards. In the author's tests it finished 20 of 20 multi-step coding tasks with zero malformed tool calls and serves OpenAI- and Anthropic-compatible endpoints. Reported result (as claimed by the author): About 100 tokens/sec generation in the author's benchmark; 7 points on Show HN. Source: https://news.ycombinator.com/item?id=50118153 Tools they used: Qwen Category: Dev & ops Please: 1. Ask me 3 short questions about my situation (accounts I have, budget, what I want the output to look like). 2. Propose a step-by-step plan using these tools or cheaper alternatives, with the time and cost of each step. Use official pricing pages; say "check pricing" if unsure. 3. Then do the setup with me one step at a time, waiting for my OK before anything that sends messages, posts publicly, spends money or deletes data. 4. Finish with a checklist to run it again next time.
You are helping me set up this AI use case end to end in Grok Bot. Goal: Run a 125B MoE coding-agent model on two 16 GB gaming GPUs What the original author did: capy-flash runs Qwen3.8-Flash-Next (125B MoE) in NVFP4 on two RTX 5070 Ti cards. In the author's tests it finished 20 of 20 multi-step coding tasks with zero malformed tool calls and serves OpenAI- and Anthropic-compatible endpoints. Reported result (as claimed by the author): About 100 tokens/sec generation in the author's benchmark; 7 points on Show HN. Source: https://news.ycombinator.com/item?id=50118153 Tools they used: Qwen Category: Dev & ops Please: 1. Ask me 3 short questions about my situation (accounts I have, budget, what I want the output to look like). 2. Propose a step-by-step plan using these tools or cheaper alternatives, with the time and cost of each step. Use official pricing pages; say "check pricing" if unsure. 3. Then do the setup with me one step at a time, waiting for my OK before anything that sends messages, posts publicly, spends money or deletes data. 4. Finish with a checklist to run it again next time.
The Daily 5
Get five new use cases tomorrow morning
Free. Real sources. Unsubscribe any time.
Related use cases
Make Codex pick a cheaper model by task effort
The author wrote a Codex skill that swaps models by effort to save usage, then found Codex running a subagent on a local Qwen model through Ollama.
Put one coding agent in charge of the others
Running Claude Code, Codex, Grok and Cursor in four terminals meant copying work between them by hand, so a former café owner and trader built Hivelane: one agent leads and the rest are on call.
Get a phone ping when your coding agent is blocked and answer over SSH
During paternity leave a developer built pocketty, an iPhone and iPad SSH terminal that knows whether each agent pane is working, needs you or done, so they could keep agents moving from the couch over Tailscale.
Run the pi coding agent in sandboxes on your own server
Built pi pod to run pi coding-agent sessions in isolated pods on a server you control, with composable environments.