Thanks to SSD streaming I was able to try DwarfStart on my M3 Max 64GB. I let DeepSeek Flash analyze a ~50K LOC codebase. Then I asked Opus to double-check and review the findings. I was impressed that 5 of the 6 findings reported by DeepSeek were accurate.
Today I had an harder than usual question for my local model (security). With SSD streaming now DwarfStar can run DeepSeek v4 PRO at 4.15 t/s, and this was more than enough to get a detailed reply. I already feel "safer" than before in my AI future. M5 max 128GB, model 433GB.
Jun 18, 2026 · 8:08 AM UTC
I meant "correct and accurate" at the level of any other SOTA model. The only incorrect finding was about a race condition that Opus wasn't able to reproduce, still a possible issue to pay attention to.
@antirez , thanks for this project 🙏
Useful datapoint. The next thing I’d want is a replay pack around the 6 findings: finding IDs, repro tests, prompt/model hashes, false-positive/triage time, and which issues survive a clean checkout. That’s what turns “LLM code review worked” into an eval harness.