MiniMax open-sources M3.1 Flash weights, claiming that its Agent task costs are reduced by 70%and generation speed is close to 3 times.
Article
MiniMax M3.1 Flash open-sourced: 428000000000 parameters,Agent task costs reduced by 70%Dongcha Beating AI Breaking News: MiniMax today open-sourced the M3.1 Flash model weights, which are now available on Hugging Face and ModelScope. Space Bunny, which was previously launched anonymously and reportedly processed more than 200000000000000 Token, is an early version of M3.1 Flash.
The model uses a MoE architecture, with total parameters of 428000000000, but only approximately 23000000000 are activated for each generation. The new version fully adopts sparse attention and adds quantization training and DSpark speculative decoding. According to the official statement,Agent task costs are reduced by 70%, and generation speed is close to 3 times.
In terms of performance, DeepSWE v1.1 scored 73.7%, lower than DeepSeek V4.1 Flash-1017's 77%. Independent evaluation organization Vals AI found that its coding capabilities have improved significantly over the previous generation, but complex terminal tasks remain a weakness.
The weights are licensed under the MiniMax Community License. Commercial use requires attribution of the model source, and prior written authorization is also required if the annual revenue of related products or services exceeds 20000000 USD.
Original link https://m.theblockbeats.info/flash/373302
Key points
01
MiniMax today open-sourced M3.1 Flash model weights, now available on Hugging Face and ModelScope.
02
The report describes the anonymously launched Space Bunny as an M3.1 early version of Flash, claiming that its cumulative processing volume exceeded 200000000000000 Token。
03
The model uses a MoE architecture, with total parameters of 428000000000, and activates only approximately for each generation 23000000000。
04
The new version uses sparse attention and adds quantization training and DSpark speculative decoding.
05
According to the official statement,Agent task costs are reduced by 70%, and generation speed is close to 3 times.
06
DeepSWE v1.1 scored 73.7%, lower than DeepSeek V4.1 Flash-1017's 77%。
07
Vals AI found that its coding capabilities have improved significantly over the previous generation, but complex terminal tasks remain a weakness.
08
The weights are licensed under the MiniMax Community License; commercial use requires attribution of the model source, and if the annual revenue of related products or services exceeds 20000000 USD, prior written authorization is also required.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
This is a model with open model weights. Although its total parameter count is 428000000000, the report states that approximately are activated for each generation 23000000000; the official statement also claims improvements in cost and speed. Its commercial use is not unconditional: attribution requirements and authorization requirements tied to revenue thresholds must be observed.
Why it matters to readers
If the cost and speed data can be replicated, developers will have another model option for Agent tasks; however, performance tests show that it still falls short on complex terminal tasks, and commercial deployment is also subject to licensing terms.
Beginners can focus on whether the model weights are easy to use, actual operating costs and speed, and whether commercial use requires source attribution or prior written authorization.
Risks and unknowns
The claimed cost reduction 70%and speed close to 3 times are official claims; the evidence does not provide details of independent replication testing.
The claim that Space Bunny's cumulative processing volume exceeded 200000000000000 Token is qualified as “reportedly” and has not been further confirmed by the evidence.
The evidence does not provide the methodology, samples, or complete results of the Vals AI evaluation.
The evidence does not show the full text of the MiniMax Community License, and the specific scope of applicability of the licensing terms still needs to be verified.
Related Developments
Loading event timeline…
Related concepts
Model weights
This term is not in the glossary yet. Browse related concepts in the glossary.
