Space Bunny (M3.1 Flash anonymous version) hands-on: strong benchmarks, weaker than GLM 5.3 Flash in OpenCode
Hands-on with Space Bunny/M3.1 Flash: strong paper scores (74% DeepSWE) but weaker than GLM 5.3 Flash for everyday coding in OpenCode, verbose and context-hungry, more implementation mistakes than expected; paid users report better results than free-endpoint users.
i’ve been testing Space Bunny / MiniMax M3.1 Flash, and the benchmarks look better than the actual experience so far: 1. on paper it’s strong: 74% on DeepSWE, 62% on Automation-Bench and 91% on BabyVision. 2. in OpenCode, it still feels weaker than GLM 5.3 Flash to me on normal coding work. 3. the model is also pretty verbose. it burns through context faster than i expected, which makes the cheap API pricing look less impressive per task. 4. i’ve seen more implementation mistakes than i expected from a model with these benchmark scores. DeepSeek still feels more reliable for actually pushing code. 5. the interesting part is that subscription/API users seem more positive than people using the free endpoint. the model may be good. the free Space Bunny endpoint might not be showing its best version.
This post quotes another post: x.com/opencode/status/2112156205938626954