MiniMax Introduces M3.1 Flash to Lower AI Agent Costs by 70%
MiniMax has released M3.1 Flash, a 428-billion-parameter model that lowers agent operational expenses by 70% and reaches speeds of up to 200 tokens per second.
Chinese artificial intelligence developer MiniMax Group Inc. has unveiled M3.1 Flash, an updated iteration of its M3 foundation model designed specifically to lower the cost and latency of running autonomous AI agents. Originally introduced in June, the M3 architecture now features aggressive inference-time sparsity that cuts agent execution costs by 70% compared to its predecessor, while boosting generation speeds to nearly three times that of the earlier build.
Running complex agentic loops often requires dozens of sequential model calls to complete a single user objective, making inference overhead and latency primary bottlenecks for autonomous workflows. According to MiniMax, M3.1 Flash addresses this friction directly, reaching everyday peak generation speeds exceeding 200 tokens per second. Despite these performance gains, the underlying footprint remains largely consistent with M3, with the preview release documentation listing a total of 428 billion parameters.
Multimodal Capabilities and Configurable Reasoning
M3.1 Flash processes prompts across text, image, and video modalities, supporting context windows spanning up to 1 million tokens. Unlike models that separate standard generation from structured contemplation, M3.1 Flash keeps reasoning enabled across every query. Instead of turning it off, developers can manage latency and compute depth by selecting one of five reasoning effort levels, ranging from Low to Max.
The model's architecture also clarifies the origin of "Space Bunny," an uncredited preview model that appeared on the OpenRouter routing platform in late September. Community members initially tied the release to MiniMax after examining its tokenizer behavior. MiniMax confirmed that Space Bunny was an early test build of M3.1 Flash, noting that the unnamed release processed more than 200 trillion tokens before its formal branding was announced.
Performance Across Specialized Benchmarks
To evaluate the system across practical agent workflows, MiniMax subjected M3.1 Flash to several technical evaluations. On the OSWorld V2 computer-use benchmark, the model posted a score of 61.7, which the company claims is the highest result recorded by a Chinese developer. In financial automation, it achieved 76.2 on the BankerToolBench evaluation.
For security analysis and technical infrastructure, MiniMax highlighted an 82% score on CyberGym, a benchmark designed by researchers at the University of California, Berkeley to measure how effectively models can replicate real-world software vulnerabilities. MiniMax indicated that these capabilities enable the model to handle tasks like automated code auditing, penetration testing, and malware inspection. On general software engineering, the system scored 73.7% on the DeepSWE v1.1 evaluation, trailing slightly behind Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co. Ltd.'s updated V4.1-Flash, which achieved 77% on the same test. Neither set of benchmark results has undergone independent verification.
What it means for developers
The arrival of M3.1 Flash introduces direct price and efficiency competition in high-volume inference markets. The model enters the market with a pricing structure of 15 cents per million input tokens, 60 cents per million output tokens, and 1.5 cents per million cached input tokens. This matches the off-peak input rate of DeepSeek's V4.1-Flash, a rival offering that previously triggered a decline in MiniMax's Hong Kong-listed shares following its September debut after MiniMax's $619 million initial public offering in January.
For teams building multi-step agent systems, the combination of 1-million-token contexts, sub-dollar output pricing, and configurable reasoning allows for broader experimentation with iterative debugging and automated tools without incurring steep API bills. Developers looking to compare and integrate such systems alongside alternative architectures can try top AI models cheaply through one API at https://apixoai.online as the broader landscape shifts toward cheaper, high-speed inference.
Access to M3.1 Flash is currently available through the MiniMax API, the company's M Plan subscription, and its MiniMax Code desktop client. MiniMax has also indicated that it intends to release open weights for Flash at a later date, alongside a more capable Pro edition.
Source: MiniMax launches M3.1 Flash model, says it cuts AI agent costs by 70% — SiliconANGLE AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
