Nine models, five coding tasks on one 16GB RX 9060 XT: deepseek-v4.1-flash hits 5/5

Ok_Substance2327 · reddit · 2026-10-18

A Redditor benchmarked nine models on five coding tasks using a single RX 9060 XT (16GB). Each run gets a spec file in an empty directory, must produce solution.py and tests, and is graded by an unseen checker with fixed cases, randomized cases against a brute-force oracle, and a performance gate; cells cap at 1,750 seconds.

Tasks: CSV quoting rules, arithmetic evaluator with pinned exceptions, POSIX-ish shell tokenizer, 200k-interval scheduling with two tie-breaks (under 3s), pattern matching with 20 stars against 4,000 chars (under 10s).

Results

deepseek-v4.1-flash (cloud): 5/5 in 236s and 43k tokens
Qwen3.8-27B UD-IQ4XS + MTP (kvstream fork): 5/5 but 94 minutes and 185k tokens; passed c5 with a Thompson NFA
Qwen3.8-27B IQ3XXS, Swift-1.5-Qwen3.8-27B, Ornith-1.5-35B-A3B, Qwen3.6-35B-A3B: 4/5 each
Nemotron-3.5-Bolt-74B-A9B: 1/5, reporting nonexistent files on multiple cells
LFM2.5-24B-A2B: 0/5

Key observations: c1-c3 just sort models into "writes code and iterates" vs not; every c4 pass used sort + bisect DP, failing on tie-breaks and duplicate-heavy input; c5 split cleanly—passes came from automata (NFA / big-integer bitset), failures from parse-then-match or Python's re backtracking on star piles. Qwen3.6-35B failed c4's gate at 3.03s vs 3.0s, with re-run variance of 2.97-3.23s.

Original post →

Nine models, five coding tasks on one 16GB RX 9060 XT: deepseek-v4.1-flash hits 5/5

More from coding & agent