Datasets:
Gemma 4-E4B: fastest portscout ever (30s), logpulse PARTIAL
added 2 data points for Gemma 4-E4B (Q4_K_M GGUF, Pi agent, RTX 4060 Ti 8GB). same benchmark tasks as existing rows.
results
| task | status | wall clock | tool calls | self-fixes |
|---|---|---|---|---|
| portscout (easy) | PASS | 30s | 4 | 1 |
| logpulse (hard) | PARTIAL | 60s | 4 | 0 |
portscout (PASS, 30 seconds)
fastest completion of any model tested. 4 tool calls: mkdir, write (88-line script), bash fail (python not found), bash retry (python3). code correct on first write: ThreadPoolExecutor with 50 workers, 0.5s timeout per port, proper reporting. 1 self-fix (python→python3).
logpulse (PARTIAL, 60 seconds)
all 3 files created correctly in 60 seconds:
logpulse.py(131 lines): argparse, deque-based 60-second rolling word window, Counter for top 5, regex alerts, tail-f file watchinggenerate_logs.py(52 lines): structured log generator with CRITICAL triggers every 15 linesREADME.md: clean documentation
code verified working externally (regex alerts fire correctly, word counting works, line processing correct).
failure mode: the model did not self-test. it wrote all files, then told the user to test manually in two separate terminals. the prompt explicitly asked to "test it by generating a fake log file with a second script."
this is different from OmniCoder and Qwopus which tried to test and got stuck on the blocking tail-f command. E4B recognised the trap and avoided it, but gpt-oss-20b solved it properly by backgrounding the generator and testing with a timeout.
comparison
| model | portscout | logpulse | total time |
|---|---|---|---|
| Gemma 4 E4B | PASS (30s) | PARTIAL (60s) | ~90s |
| gpt-oss-20b | PASS (~3 min) | PASS (~5 min) | ~8 min |
| OmniCoder 9B | PASS (<1 min) | PARTIAL (stuck) | — |
| Qwopus 9B | partial (~8 min) | FAILED | — |
key takeaway
E4B is the fastest code generator on this benchmark: 30s portscout is 6x faster than gpt-oss-20b. code quality is excellent (6/6 on the quality battery, clean first-shot code). but ~4B effective params can't orchestrate multi-step workflows involving blocking commands. gpt-oss-20b remains the only model completing both tasks on 8GB VRAM.