{
  "schema": "genie-recipe.v1",
  "id": "sovereign-coding-room",
  "name": "Sovereign Coding Room — Qwen3 Coder",
  "stage": "benchmarked",
  "commercial": false,
  "verified_at": "2026-08-03T00:35:00Z",
  "model": {
    "id": "Qwen/Qwen3-Coder-30B-A3B-Instruct",
    "checkpoint": "official BF16 weights",
    "license": "Apache-2.0"
  },
  "runtime": {
    "engine": "vLLM nightly",
    "context_tokens_tested": 8192,
    "max_sequences": 8,
    "gpu": "1x NVIDIA A100 80 GB",
    "gpu_hour_usd_measured": 1.39
  },
  "evidence": {
    "ready_seconds": 139.57,
    "single_request_output_tok_s": 109.78,
    "batch_2_aggregate_output_tok_s": 159.09,
    "batch_4_aggregate_output_tok_s": 221.12,
    "batch_8_aggregate_output_tok_s": 517.04,
    "batch_8_median_latency_s": 2.97,
    "coding_first_pass": "4/5",
    "native_tool_call_gate": "passed",
    "teardown_gate": "passed"
  },
  "modeled_cost": {
    "compute_per_million_output_tokens_usd": 0.7468,
    "minimum_session_price_per_hour_for_50pct_gpu_margin_usd": 2.78,
    "versus_deepseek_throughput_x": 3.05,
    "versus_deepseek_compute_cost_advantage_x": 16.19,
    "versus_deepseek_compute_cost_reduction_pct": 93.82
  },
  "required_policy": {
    "concurrency": 8,
    "compile_and_test": true,
    "maximum_repair_attempts": 1,
    "idle_teardown_minutes": 5,
    "commercial_lifecycle_canaries_required": 3
  },
  "known_boundaries": [
    "The five-task coding suite is a smoke test, not HumanEval or SWE-bench.",
    "One coding task failed, so every patch must compile and test before delivery.",
    "Evidence-marketing and filing-citation tasks failed manual review; this runtime is promoted for coding only.",
    "Payment, delivery, expiry, and teardown canaries are not yet complete, so this is capacity review rather than checkout inventory."
  ],
  "heavyweight_comparator": {
    "model": "deepseek-ai/DeepSeek-V4-Flash-0731",
    "gpu": "1x NVIDIA B300 288 GB",
    "batch_8_aggregate_output_tok_s": 169.74,
    "coding_first_pass": "4/5",
    "compute_per_million_output_tokens_usd": 12.09,
    "raw_cold_start_minutes": [21.53, 29.30]
  },
  "files": {
    "launch": "/recipes/qwen3-coder-30b-a3b/serve.sh",
    "documentation": "/recipes/qwen3-coder-30b-a3b/README.md"
  }
}