Qwen3.8-27B on RTX 3090 with OpenCode and measured coding results

Description and use

The RTX 3090 runs Qwen3.8‑27B locally, achieving 20.9 tokens/s on 120 K input and 0.63 s first‑token latency for cached continuations. Four small Python projects were tested; three passed all independent checks and two completed within the eight‑minute deadline.

Problem

Determine if a legacy RTX 3090 can run a large local coding model (Qwen3.8‑27B Q4_K_M) and support a practical coding workflow via OpenCode.

Outcome

The RTX 3090 successfully hosts Qwen3.8‑27B, achieving 20.9 tokens/s on 120 K input and 0.63 s first‑token latency for cached continuations. Four small Python projects were tested; three passed all independent checks and two completed within the eight‑minute deadline. The build‑planner task failed due to exhausting the 8,192‑token output budget.

Workflow

• Load Qwen3.8‑27B Q4_K_M weights (~16.8 GB) into llama.cpp on the 3090.
• Configure llama-server to expose a local API at 127.0.0.1:8080/v1.
• Connect OpenCode to the local API; OpenCode sends messages and tool definitions.
• llama-server runs inference on the GPU and returns text or tool calls.
• OpenCode executes returned tool calls in its workspace and feeds results back to the model.
• Repeat until the model produces a patch or the output budget is exhausted.
• Run independent test suites outside the workspace to validate the patch.
• If successful, hand off the patch for final review; otherwise, retry or abort.

Evidence

We loaded Qwen3.8‑27B Q4_K_M into the 3090's 24GB of VRAM, served it through llama.cpp, and connected OpenCode.

Source

With short input, generation reached about 36.4 tokens/s. At roughly 120K input tokens, it fell to 20.9 tokens/s, and a fresh request took about three minutes to produce its first token.

Source

Three candidates passed all predefined independent checks; two completed editing, testing, and final handoff within the deadline.

Source