Methodology
Smoke24 is a 24-task slice of Terminal-Bench 2.0 designed for fast, repeated comparison of coding-agent behavior.
Rationale
The subset was created to answer a practical local iteration problem: can a new model, quantization recipe, context setting, or serving setup be assessed on an RTX 3090-class machine within a couple of hours?
The target is a benchmark that is small enough to run frequently on local hardware, while still exercising real Terminal-Bench agentic tasks. It is meant to provide an early signal on model quality, runtime behavior, token use, parser stability, and failure modes before spending time on broader or more expensive benchmark passes.
The local rows in this project are interpreted under one serving boundary:
- vLLM or LiteLLM residency serving
- RTX 3090-class local deployment context
- Terminus-2 agent harness
- 30 minute task timeout
- 32 CPU / 48 GiB sandbox shape
The dashboard emphasizes three classes of signal:
- Quality: verifier success rate on the same 24 tasks.
- Local serving cost: LLM API minutes, wall-clock minutes, generated tokens per solved task.
- Stability: historical Smoke24 pass variance, parser-warning counts, task-level pass/fail matrix.
External References
External public Terminal-Bench rows are reconstructed as equivalent Smoke24 success references. They are not hardware-comparable for local cost or wall-clock metrics, so they should appear as reference lines or success-only tables rather than local-cost scatter points.
Variance
Smoke24 is intentionally small. One-task deltas can be meaningful but should not be over-read without historical pass variance. The report keeps score, token, LLM-time, observed decode-speed, exception, and parser-warning variance in separate aggregate files.