# Qwen3 Coder 30B-A3B — Sovereign Coding Room

Status: **BENCHMARKED / capacity review**. It is not checkout inventory yet.

This is now Genie's primary sovereign coding runtime. On one RunPod A100 80 GB
at $1.39/hour, the official BF16 checkpoint reached readiness in 139.57 seconds,
passed 4/5 executable coding tasks, and sustained 517.04 aggregate output tok/s
at eight-way concurrency. Provider teardown passed.

## Why it replaced DeepSeek as the default

The same five-task coding gate gave both models 4/5. Qwen3 Coder delivered:

- 3.05x the eight-way throughput: 517.04 versus 169.74 tok/s.
- 16.19x lower measured GPU token cost: $0.7468 versus $12.09 per million
  output tokens.
- 139.57-second readiness versus DeepSeek's 21.53–29.30-minute raw cold start.

DeepSeek remains a heavyweight comparator for explicit requests; it is no
longer Genie's primary coding runtime.

## Intended use

- Confidential repositories and bounded private work sessions.
- Four to eight independent implementation, refactor, or test tasks.
- Test-gated patch batches where every output is compiled and executed.

Do not use this evidence to sell the model as a general research, marketing, or
filing-analysis agent. Manual review rejected its unsupported marketing claim
and missing filing citations.

## Operating policy

Run at no more than eight concurrent tasks. Compile and test every patch, allow
one repair attempt, and then escalate. Tear the GPU down after five idle
minutes. The $2.78/hour figure is only the GPU-cost floor for 50% gross margin;
quotes must also cover startup, orchestration, transfer, support, and failed
capacity.

Before commercial checkout is enabled, Genie requires three payment,
provisioning, delivery, expiry, and teardown canaries plus named demand.

Files:

- Evidence manifest: `/recipes/qwen3-coder-30b-a3b/recipe.json`
- Launch recipe: `/recipes/qwen3-coder-30b-a3b/serve.sh`