DGC has no time limit — it's built for local models, and it works a hard problem until it's done. Given that room, on a single 27B on your own machine, it solves all but one of the benchmark's 90 problems across six languages.
A coding agent's score is mostly its harness — how it reads,
edits, runs tests and recovers — not the model underneath. So we fixed the model to one 27B anyone can run,
and measured DGC on the standard Aider polyglot set: 90 exercises, six languages, scored by each one's real
test suite, pass@2. DGC's only miss is rust/doubly-linked-list — unsafe-pointer surgery the
field-leading harness fails too. The genuine hard tail, not a harness gap.
Leaderboards run a tight 600-second-per-problem cap and reward pure sprint speed. DGC optimizes for what the stopwatch can't see: it's built to be pointed at your real repository — permission controls, sandboxing and credential redaction on, running entirely on your own machine, with no clock forcing a slow local model to give up early. Playing on the clock's terms anyway, on the same model, DGC lands level with OpenCode and a couple of problems behind the leaders.
| Harness | Pass@2 · 600 s cap | Score |
|---|---|---|
pi | 98.9% | |
Codex CLI | 97.8% | |
DGCthis is us | 95.6% | |
OpenCode | 95.6% | |
goose | 94.4% | |
Aider | 77.8% |
Lift the cap to the time a local model actually needs — DGC's real design point — and it solves 98.9%. The 600 s sprint is the competitors' home turf; the marathon on your own hardware is ours.
All agents drive qwen3.8:27b-bf16 through its OpenAI-compatible endpoint. Correctness is hardware-independent.
15 exercises each in Python, Go, Rust, JavaScript, Java and C++ — a deterministic slice of the Aider polyglot set, pinned to one dataset commit.
Round 1 implements and self-tests; if the real suite fails, round 2 gets the failure and retries. Scored only by the exercise's own tests.
No time cap and DGC's shipped iteration budget — exactly what a user gets. The 600 s figures use the leaderboard-standard cap for a like-for-like comparison.
The whole harness ships with DGC. Point it at any OpenAI-compatible model and the same numbers fall out — for DGC or any of the other agents.
# get the harness (bundled with DGC) git clone https://github.com/OpenPeach-ai/dgc && cd dgc/bench bash install_harnesses.sh # DGC + the other agents + toolchains # run DGC on the 90-problem subset python3 run_bench.py --engine dgc --model qwen3.8:27b-bf16 \ --base-url http://localhost:11434/v1 --limit 15 --rounds 2
Coming next: SWE-bench — real GitHub issues from Django, SymPy, scikit-learn and more, on the same local model. We'll add the results here.