The performance and reputation of MiniMax M3.1 Flash vary greatly

---

First, lock the window. The official blog of MiniMax M3.1 Flash was released on October 19, 2026, and a discussion post appeared on Hacker News on the same day, with story ID 50154540. At the time of snapshot, there were 324 upvotes and 140 comments. Calculated over a 48-hour window, the initial release date is October 19, 2026. This article was written on the same day and is still within the window. The required context should also be clearly stated upfront: the official benchmark results differ from community evaluations, which is the main focus of this article. However, since the weights have not been released, ordinary users and third parties cannot replicate these numbers independently. Therefore, the article will not take official data as definitive conclusions but will list the sources for each item and maintain room for doubt.

What Happened

On October 19, 2026, MiniMax released M3.1 Flash, a fast model designed for agentic tasks, coding, computer use, and cybersecurity scenarios. The core selling point offered by the official team is quite impressive:推理 speed is nearly three times faster than previous versions, usage cost is 70% lower, it reaches leading levels in multiple benchmarks, and it has already processed over 200 trillion tokens when running anonymously as Space Bunny.

In terms of product lines, MiniMax offers three options: API, M Plan subscription, and the MiniMax Code desktop client. The API is pay-as-you-go for developers, M Plan is a monthly subscription for frequent users, and MiniMax Code directly competes with AI coding tools like Cursor and Windsurf. This “model + application + subscription” combination indicates that MiniMax aims not only to sell tokens but also to keep users within their own workflows. For observers, determining whether M3.1 Flash is successful requires looking at not only the model leaderboard but also the retention rate of MiniMax Code and the conversion rate of M Plan.

In terms of pricing, M3.1 Flash is priced at $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens. The official performance data includes: 76.2 points on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 points on OSWorld V2, 29.6 points on Agents’ Last Exam, and 82% on CyberGym. It also claims to exceed 200 TPS in output speed, with a median P50 of over 170 TPS. These numbers point to one conclusion: this is a fast and affordable product that can compete with top models in certain scenarios.

In terms of availability, the API, M Plan, and MiniMax Code have been launched, while the weights will be released later. This means that currently, the main thing that can be directly verified is the API experience, and to reproduce the official benchmark, we need to wait until the weights are available.

From the release strategy, MiniMax chose to first conduct anonymous tests with Space Bunny before unveiling its product. The fact that 200 trillion tokens were processed during the anonymous testing period indicates that the test scale is large enough, and it also shows that the manufacturer has collected a large amount of real feedback before release. The problem is that this feedback is solely held by the manufacturer, and external parties cannot see the original logs, making it impossible to determine the exact relationship between “200 trillion” and “quality.” For the technical community, this kind of “internal testing” can only serve as circumstantial evidence and cannot replace publicly verifiable benchmarks.

The Gap Between Benchmarks and Public Opinion

Official data seems abundant, but early feedback in the HN discussion forum has not been fully accepted. This gap is mainly reflected in three aspects.

First is direct comparison. Commenter johnnyApplePRNG mentioned that M3.1 Flash’s score on DeepSWE is lower than DeepSeek V4.1 Flash 1017. Unless MiniMax can reduce its price to half or less, it will be difficult to make it attractive. The implication of this comment is clear: the benchmark lead is conditional; once compared with similar prices or in the same ecosystem, the advantage will narrow.

Second is the public opinion during the anonymous testing period. Commenter walrus01 pointed out that the scores of DeepSWE and TerminalBench 4.0 do not fully match the Space Bunny experience measured by OpenCode users before, with obvious errors at times. This means that beyond the official testing, early users have already accumulated a group of real memories that “it also makes basic mistakes.”

Third is the verification conditions. Commenter terhechte noticed that since the weights have not been released, it can only be inferred based on the official statement that this might be a model with about 150B parameters, perhaps capable of running on consumer-grade hardware. Before the weights are available, all benchmarks can only be considered as results provided unilaterally by the manufacturer, and third parties cannot replicate them independently.

Putting these three points together, the core issue is not that “official data is definitely fake,” but rather there is an explanation gap between the “highly praised numbers” given by the official side and the “actual performance perceivable by early users.” Whether the model is truly that fast, accurate, or inexpensive can only be determined after more third-party evaluations and the release of the weights.

The Official Statement Does Not Match

This section is not about accusations, but rather about putting together what the official side has said and hasn’t fully stated.

The core narrative on the official blog is: faster, cheaper, leading in multiple benchmarks. But the term “leading” does not provide a control group. Is it compared to MiniMax’s previous generation? Or to DeepSeek Flash of the same price? Or to flagship models at GPT-4 level? Different control groups would give completely different meanings to “leading.”

Furthermore, the official benchmarks provide absolute scores rather than head-to-head comparisons with competitors under the same conditions. Numbers like 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, and 61.7% on OSWorld V2 are high individually, but it’s difficult to determine their actual meaning without knowing which version, prompt, or evaluation script they were run under. The reproducibility of benchmarks depends not only on whether the models are publicly available but also on whether the testing environment is transparent.

The official statement also mentions that during the anonymous operation of Space Bunny, over 200 trillion tokens were processed. This figure spreads quickly, but it describes the total number of calls made, not the model’s actual capability. How many times a model is called has no direct relationship with the quality of each response.

Another unclear statement is about the measurement condition of “almost 3 times faster than the previous generation.” For the same model, TPS can vary several times depending on input length, concurrency, and hardware configuration. The official claimed peak TPS of over 200 and P50 median of 170 TPS are hard to verify without a description of the testing environment.

Finally, the fact that the weights have not been released makes the official benchmarks unverifiable independently. This is no small issue for the technical community. Reproducibility is the foundation of the credibility of benchmarks; without it, even impressive numbers can only be placed in the “pending verification” folder.

What were the comments like?

Among 140 comments on HN, opinions roughly fell into four categories, all treated as quoted.

The first category is the comparison faction. In addition to johnnyApplePRNG comparing DeepSeek V4.1 Flash 1017, swiftcoder was curious about whether the price/performance ratio of M3.1 Flash could truly match DeepSeek Flash. The core idea of these comments is that official benchmarks should be viewed in the context of competitors, rather than claiming “leadership” in isolation.

The second category is the skeptical faction. walrus01 compared the official scores with user anecdotes from the anonymous period of Space Bunny, noting a gap between the two. terhechte pointed out that since the weights were not released, even the model size could only be guessed. These two comments together point to the same issue: before independent verification, the credibility of the official benchmarks can only be discounted.

The third category is the information organizer faction. Some commenters focused on whether the model could run on consumer-grade hardware and whether the pricing was indeed 70% lower than competitors, as claimed by the official. The answers to these questions are currently incomplete, but they reflect two key concerns of the community: whether it can run locally and whether the actual cost is high. For developers, these two indicators are more direct than any leaderboard score.

The fourth category is the meta-discussion faction. theanonymousone noted that this was duplicate posting, with multiple posts on the same topic, indicating that MiniMax’s release on HN indeed caused a stir, but it also raised doubts about whether the discussion heat was exaggerated by marketing. Duplicate posting does not prove product quality, but it reminds us: heat and quality are two different things and should not be confused.

Three Things Ordinary Users Can Do

First, question the official benchmarks before attempting to reproduce them with third-party tools. Until weights and independent evaluations are available, use official data as a reference rather than a final conclusion. Pay special attention to relative values like “leading,” “3 times faster,” and “70% lower cost”; their comparison groups often lie in footnotes.

Second, if you want to test the model, first run it under low-traffic conditions in your own environment. A high score on public benchmarks doesn’t mean it will work well in your code base, documentation structure, or business processes. MiniMax Code and the API are now available; try them with real tasks before deciding whether to migrate your main workflow.

Third, consider M3.1 Flash as a “candidate” rather than the “only solution.” There are multiple fast models at this level, including DeepSeek Flash, GPT-4o mini, and Claude Haiku, all of which are continuously updated. MiniMax’s pricing and performance are indeed competitive, but the market landscape changes rapidly. It’s safer to make long-term investments after two weeks to a month of community feedback.

One final note: For small businesses and independent developers, “price” is often the most tempting factor when choosing a model, but real costs are often hidden. Costs for migration integration, prompt engineering, error handling, and compliance checks can far exceed the token bill itself. Switching models frequently just to save a few dollars may not be worth it in the end. Consider M3.1 Flash as a new option worth observing, not a reason to immediately replace your existing workflow.

From a broader perspective, MiniMax’s release is a reflection of domestic large model companies competing in price, speed, and benchmarks. As companies tout their “faster, cheaper” models, users should be wary of treating individual metrics as an overall experience. Model selection isn’t like lottery; you can’t rely solely on leaderboards.

Conclusion

Statement of limits: This article doesn’t accuse MiniMax of anything, nor does it deny the technical value of M3.1 Flash. The official benchmarks and performance data come from MiniMax’s blog and have been labeled accordingly; comparisons and doubts in the HN comments are shared from the community and also attributed individually. Feedback from users during Space Bunny’s anonymous testing phase is anecdotal and cannot prove the model’s quality, but it can explain the “gap” between models. The fact that weights haven’t been released and cannot be reproduced independently by third parties runs through the entire article.

Verified facts include: MiniMax’s official blog published M3.1 Flash on October 19, 2026; HN story 50154540 was submitted on the same day, with a snapshot of 324▲/140💬; the pricing is $0.15/$0.60/$0.015 per million tokens; APIs, M Plan, and MiniMax Code are available, with weights to be released later; official benchmarks include BankerToolBench 76.2, DeepSWE v1.1 73.7%, OSWorld V2 61.7%, ALE 29.6%, CyberGym 82%, and output speed of 200+ TPS.

There are two key points to monitor: First, whether the community can reproduce the official benchmark after the weights are officially released; second, the cross-compare results of third-party evaluation agencies or developers on models comparable to DeepSeek Flash, GPT-4o mini, etc. Any new evidence sufficient to change the current judgment will prompt the release of v2.

Reference Sources

• MiniMax Official Blog 2026-10-19「MiniMax M3.1 Flash」: Model release, performance data, pricing, and availability sources, https://www.minimax.io/blog/minimax-m31f
• HN story 50154540: Submitted on 2026-10-19T16:39:09Z, title「MiniMax M3.1 Flash», snapshot 324▲/140💬; Commenters swiftcoder/terhechte/johnnyApplePRNG/walrus01/theanonymousone are their own identifications, with details attributed accordingly.
• Fact package /Users/kael/xhs-content/drafts/2026-10-20-facts-packages.json’s「hot」entry: The only fact source for this article. The entire writing process was conducted offline without rechecking online.