After using MiniMax M3.1 Flash for several hours, J A Z I I considered it a clear improvement over M3, with particularly strong performance in Computer Use and Three.js. However, it was prone to hallucinating and carrying out unrequested actions at xhigh or max, so the author positioned it as a sub-agent for simple tasks rather than a daily driver.
Tasks this report can inform: Computer Use, Three.js/frontend coding, and trying the model as a sub-agent for simple tasks.
Areas where extrapolation is not appropriate: General coding success rates, stability, speed rankings, fair quantitative comparisons with DeepSeek, or reliability in any production task.
Model version covered: The author's original text uses minimax m3.1 flash; no more specific model snapshot or API ID was disclosed.
Test environment or client: Not specified. The author only said the model was used continuously for several hours, and mentioned Computer Use, Three.js, and a comparison with DeepSeek using the same prompt.
Reasoning levels and parameters: The author mentioned high, xhigh, and max; temperature, context, tool permissions, concurrency, model provider, and other parameters were not specified.
This is a personal preliminary-use report, not a controlled evaluation under a standardized harness. The author said they had used the model for several hours and observed Computer Use tasks, Three.js tasks, and other outputs. No complete task list, prompts, acceptance criteria, number of runs, or scoring rules were disclosed.
An addition in the same thread said that, in a test whose details were not disclosed, one MiniMax attempt produced a better result than DeepSeek. The author said DeepSeek used the same prompt and tried 3+ times, while MiniMax succeeded on the first attempt. The original post does not specify DeepSeek's full version, client, reasons for failure, or other configuration on either side, so this can only be treated as the author's informal comparison rather than a benchmark result.
The author considered M3.1 Flash a “huge improvement” over M3, and called it “pretty good” for a Flash model.
The author said some outputs were “really surprising” and that Computer Use was “really strong.”
The author said the model performed well on Three.js and, on some tasks, surpassed “the latest DeepSeek” on the first attempt; the specific tasks, model versions, and scoring criteria were not specified.
The author also noted that other outputs were far worse than expected; at xhigh or max, the model hallucinated more and began carrying out things the user had not requested.
The author said the model could reach 100+ tok/s, but overall use was still relatively slow because of hallucinations and extra behavior. No measurement protocol, input or output lengths, or statistical distribution was provided for this speed.
The author's recommendation for daily use was to treat the model as a sub-agent for simple tasks; the overall assessment was “good model, but not impressed.”
In an addition in the same thread, the author said hallucinations were less frequent and the model was faster at high; this remains a single user's observation with no repeated runs or quantitative record.
| Item | Information visible in the original |
|---|---|
| Model | minimax m3.1 flash; snapshot or API ID not disclosed |
| Usage duration | The last few hours; exact duration not specified |
| Tasks or capabilities | Computer Use, Three.js; other tasks not elaborated |
| Reasoning levels | high, xhigh, max |
| Speed | The author reported 100+ tok/s; measurement conditions not specified |
| DeepSeek comparison | Same prompt; the author said DeepSeek tried 3+ times and MiniMax succeeded once |
| Result determination | The author said MiniMax won this test; the full prompt, outputs, scoring criteria, and sample count were not disclosed |
| Client, provider, and parameters | Not specified |
The author's additions directly related to this conclusion in the same thread:
Same-prompt comparison and first-attempt success: The author said DeepSeek tried 3+ times with the same prompt, while MiniMax completed it once.
Hallucinations causing slower performance: The author said that although this is a Flash model, hallucinations made it slower.
Observation at the high reasoning level: The author said hallucinations were less frequent at high, making the model faster.
One successful attempt versus multiple DeepSeek attempts: The author again explained that MiniMax's result came from one attempt, while DeepSeek's result took 3+ attempts.
The limited conclusion supported by this record is that, in the author's several hours of personal use, M3.1 Flash was considered suitable for Computer Use, Three.js, and simple sub-agent tasks. The author also observed subjective differences between high, xhigh, and max, and considered the higher levels more prone to hallucination and instruction drift. The author reported an asymmetric one-off comparison with DeepSeek, but did not disclose enough material to reproduce it.
This does not establish M3.1 Flash's general coding ability, Computer Use success rate, actual throughput, ranking across reasoning levels, or DeepSeek's overall disadvantage. The 100+ tok/s figure lacks a defined measurement basis; “one successful attempt” and “3+ attempts” were not accompanied by the task, full prompt, failed outputs, or acceptance criteria; and the “latest DeepSeek” mentioned by the author has no specified version. This article does not present the demonstration or these opinions as a controlled evaluation.
The original post does not disclose enough to reproduce the result strictly: prompts, tasks, client, model snapshot, tool configuration, sampling parameters, sample count, and scoring rules are all missing. To reproduce it, specify the exact API IDs/snapshots for MiniMax and DeepSeek, the client, and tool permissions; publish the same Three.js or Computer Use task, full prompt, acceptance criteria, reasoning level, and number of repetitions; and record first-attempt success, hallucinations or unauthorized actions, output tokens, end-to-end latency, and token speed separately. Only after conditions are matched and runs are repeated can this personal experience be extended into a comparative result.
MiniMax M3.1 Flash
