Flash models are shifting from being cheaper alternatives to flagship models into a new contest of capability.
There was a time when the most important battle among large models was about who could push the ceiling of capability higher, but as the year progressed, the wind changed. The industry became more focused on who could turn model capabilities into productivity that developers and enterprises could genuinely afford, at a lower cost.
On October 17, DeepSeek suddenly released the V4.1 Flash-1017 update. Almost at the same time, MiniMax also brought M3.1 Flash to the stage. The latter had previously appeared in the developer community as an anonymous model called Space Bunny, and now its identity has finally been revealed.
The back-to-back arrival of these two models has given the domestic large model Flash track a new point of interest.
In the latest benchmark data, M3.1 Flash scored 74 on the DeepSWE test, second only to DeepSeek V4.1 Flash-1017's 77. In the financial Agent benchmark BankerToolBench, it scored 76.2, tied for first place with Claude Opus 5. At the same time, MiniMax cut the input and output prices of the new model to half those of the previous generation M3, and reduced cache read pricing by about 75%.
The simultaneous advance of price and capability is what makes this update worth watching.
In the past, Flash usually meant a trade-off — sacrificing some capability in exchange for faster responses and lower costs. But the report cards delivered in succession by DeepSeek and MiniMax are challenging that industry inertia.
A new strong contender enters the Flash tier
Tonight, MiniMax M3.1 Flash was officially released. What is pleasantly surprising about this update is that, as a Flash-tier model, its capabilities have improved substantially.
According to MiniMax's disclosure, compared with the previous generation M3, M3.1 Flash has made clear progress in areas such as coding, cybersecurity, and complex Agent tasks. In multiple benchmarks, it is already able to go head-to-head with flagship models, with some scores even higher.
In addition, M3.1 Flash scored 61.7 in OSWorld V2, which measures AI computer operation and cross-application task execution capabilities, and scored 29.6 in Agents' Last Exam (ALE), which evaluates performance in real professional work scenarios and complex task completion, ranking first among domestic models in both.
What is even more noteworthy is the cost of completing these tasks. M3.1 Flash's pass rate is higher than that of other flagship domestic models, while the cost of running the entire benchmark suite is lower. This means Flash models are not only starting to challenge the capabilities of flagship models, but are also attempting to complete complex professional tasks at a lower cost.
In the DeepSWE v1.1 benchmark, M3.1 Flash achieved a score of 73.7%, also notably higher than several flagship models. This further shows that in specific coding task evaluations, the capability gap between Flash and flagship models is no longer an insurmountable chasm.
Cybersecurity and complex Agent tasks are the spaces M3.1 Flash is trying to open up further.
In the CyberGym benchmark, the model scored a high 82%, putting it in a similar range to GPT-5.6 Sol's 83.6. CyberGym covers 1,507 historical vulnerabilities across 188 real software projects, mainly examining a model's ability to analyze software vulnerabilities.
In Automation-Bench v1.0.6, M3.1 Flash scored 62.2; in BankerToolBench, it scored 76.2, tied for first with Claude Opus 5. The latter focuses on multi-step planning in financial scenarios, cross-document consistency, business compliance, and complex tool invocation.
This means that in coding, financial tool invocation, and some enterprise automation tasks, lightweight models are beginning to have the ability to participate in highly complex work.
Previously, Space Bunny attracted attention in the developer community as an anonymous model. Its identity sparked outside speculation for a time, and beyond performance, speed was also one of the most frequently mentioned characteristics in developer feedback at the time. As MiniMax claimed the model, the connection between Space Bunny and M3.1 Flash gradually became clear.
Because it uses a more aggressive sparse structure, M3.1 Flash's output speed in daily usage scenarios can reach 200+ TPS, with a median (P50) steady at 170+ TPS, meaning 170 tokens generated per second.
For ordinary users, a faster model response means shorter waiting times. But for developers, speed means far more than interaction experience.
When a model begins to take on continuous tasks such as code writing, automated office work, and data processing, response speed directly affects the overall efficiency of the workflow. This is especially true in Agent scenarios, where the model needs to repeatedly understand the task, call tools, read results, and then decide the next action. The time saved in a single call may accumulate continuously across multiple rounds of execution.
Flash twin stars, with value for money maxed out on both sides
MiniMax chose this moment to unveil M3.1 Flash, and it has already formed a "Flash dual-strong pattern" with DeepSeek.
On October 17, DeepSeek updated V4.1 Flash-1017. Previously, DeepSeek had already established a distinct market perception around low-cost models. This update pushed coding capability to the front ranks, further strengthening the competitiveness of its Flash model.
MiniMax followed closely behind, and in terms of its killer features, the two share some similarities.
For example, both vendors sit in a similar price band, directly maxing out value for money.
M3.1 Flash costs 1 yuan per million tokens for input, 4 yuan for output, and 0.1 yuan for cache reads. Compared with the corresponding prices for M3 (input ≤ 512K tokens), both input and output unit prices are reduced by more than 50%, and the cache read unit price is reduced by more than 70%. Its pricing is equivalent to about half of DeepSeek V4.1 Flash-1017's peak-hour price; in cache-hit scenarios, the two are basically on par.
It should also be noted, however, that DeepSeek has differences between peak and off-peak periods, and actual costs are also affected by cache hit rates, the ratio of input to output tokens, and the number of task execution rounds. A single price cannot simply be used to infer the final cost in all scenarios.
But the trend the two have jointly set off is already quite clear.
When model capabilities are close, price differences change developers' choices; when prices fall further, applications that could not be deployed at scale in the past because costs were too high may also regain commercialization opportunities.
For enterprises, what truly matters is not how many points higher one model scores on a leaderboard, but how much money and time it takes to complete the same task, and whether repeated model calls are needed after failures.
Therefore, DeepSeek and MiniMax are more like two important options in a developer's toolbox. What they jointly promote is a rebalancing between model capability and usage cost.
In MiniMax's DeepSWE benchmark, DeepSeek V4.1 Flash-1017 scored 77 and M3.1 Flash scored 74, and the gap between them is no longer large. For developers focused mainly on code generation, code modification, and software engineering tasks, after the release of MiniMax M3.1 Flash, DeepSeek is no longer the only benchmark for value for money.
And in MiniMax's BankerToolBench evaluation, M3.1 Flash tied for first with Claude Opus 5 at 76.2. Combined with its performance on Automation-Bench, MiniMax's advantages fall more on complex task planning, tool invocation, and cross-system workflows.
Their core direction is the same: to let AI do one thing well, there can be a cheaper and more usable way.
MiniMax is back
For MiniMax itself, the emergence of M3.1 Flash carries another layer of meaning: it offers an opportunity to re-examine MiniMax's text model capabilities.
For a period of time, MiniMax's label in the market was more associated with multimodal capabilities. By comparison, competition in text models is more brutal. Reasoning, coding, and tool invocation each require long-term investment; and when model capabilities gradually converge, price, speed, and actual task completion rates become the new dividing lines.
The previous M3 did not fully dispel outside doubts about MiniMax's text model capabilities. M3.1 Flash seeks to respond to these questions with more specific results.
From the benchmark data, the new model shows clear improvement over M3, while continuing its relatively fast output speed. Its performance in coding, cybersecurity, and complex Agent tasks also gives MiniMax a chance to re-enter the core competition in text models.
For MiniMax, what is more worth watching next is whether this capability improvement can extend to a broader range of task scenarios, and whether it can continue to improve the model's reliability while maintaining its advantages in speed and price.
As for whether MiniMax can use this to re-establish its competitive position in text models, M3.1 Flash is more like a new starting point rather than a final answer. Under the new capabilities refreshed by Flash, one can look forward to the release of the MiniMax Pro model. At least judging from the benchmarks disclosed this time, the release of M3.1 Flash gives MiniMax more persuasive chips in text model competition.
The backdrop to this competition is that the large model industry is moving from capability demonstrations to efficiency competition.
When model capabilities are sufficient to handle an increasing number of real tasks, enterprises no longer care only about parameter scale and leaderboard rankings, but about the comprehensive cost of completing tasks. Cheaper tokens, faster responses, and fewer invalid calls must ultimately translate into measurable business returns.
The successive moves by DeepSeek and MiniMax mean Flash is no longer just a low-priced substitute for flagship models. It is becoming an independent competitive track: it must be smart enough, fast enough, and cheap enough.
The next round of competition among domestic large models may no longer be about who can push prices to the lowest, but about who can still do complex things well when prices are already low enough.
That is the question MiniMax needs to continue answering after returning to the table.
Comments