MiniMax Launches M3.1 Flash: High-Speed Agentic Model Slashing Costs by 70%
Chinese AI lab MiniMax has debuted M3.1 Flash, an ultra-fast foundation model engineered for agentic loops, coding workflows, and computer use, delivering 3x faster inference throughput, 70% lower operational costs, and an upcoming open-weights release.
Hong Kong-listed Chinese artificial intelligence laboratory MiniMax has formally introduced its latest frontier model iteration: 'M3.1 Flash'. Diverging from conversational models engineered primarily for single-turn dialogue, M3.1 Flash is architecturally optimized for long-horizon agentic task loops, multi-step software engineering, operating system navigation (computer use), and automated cybersecurity workflows.
In its technical disclosure, MiniMax officially confirmed that the viral unlabelled model 'Space Bunny'—which processed more than 200 trillion tokens across developer communities on OpenRouter since late September—was the early testing build of M3.1 Flash. The release intensifies competitive pressure across the lightweight model tier following DeepSeek's V4.1-Flash rollout.
Sparse Attention Dynamics: 200+ Tokens/sec and a 70% Cost Reduction
M3.1 Flash leverages an aggressive sparse attention routing mechanism, driving generation throughput to peaks above 200 tokens per second (with median output sustained around 170 tokens/sec)—delivering a near threefold inference speedup over previous generations.
This acceleration is coupled with major unit-cost optimization: MiniMax claims overall expenditures per successful agentic task dropped by 70%. Global API pricing is pegged at US$0.15 per 1 million input tokens, US$0.60 per 1 million output tokens, and US$0.015 per 1 million prompt-cached input tokens.
Benchmark Profile: DeepSWE 73.7% and OSWorld V2 Computer Use 61.7
MiniMax published extensive internal benchmark results positioning M3.1 Flash at the frontier of compact agentic models:
- DeepSWE v1.1 (Software Engineering): Recorded a 73.7% success rate resolving end-to-end repository GitHub issues and code patch generations.
- OSWorld V2 (Computer Use & GUI Navigation): Scored 61.7 navigating desktop environments, file managers, and complex browser workflows.
- CyberGym (Cybersecurity Defense): Achieved 82% success across automated vulnerability assessment and intrusion mitigation challenges.
- BankerToolBench & AutomationBench-AA: Reached 76.2 across structured financial compliance APIs and enterprise process automations.
- Agents' Last Exam: Registered a 29.6 score on one of the industry's most challenging multi-step agent reasoning evaluations.
Ecosystem Deployment & Open-Weights Commitment
M3.1 Flash is accessible across the global MiniMax API, the consumer M Plan subscription, and the MiniMax Code IDE integration. Significantly, MiniMax announced that open model weights will be released to the global developer community on October 23, 2026, accompanied by an upcoming M3.1 Pro tier.
For agentic software developers, sub-second latency is transformative. In iterative coding pipelines—reading context, proposing edits, executing test suites, parsing stack traces, and regenerating patches—eliminating intermediate latency dramatically accelerates developer productivity.
- MiniMax launches M3.1 Flash, unmasking the viral 'Space Bunny' model that processed 200T tokens on OpenRouter
- Features sparse attention delivering 200+ tokens/sec generation (near 3x faster than prior models)
- Slashes per-task agent execution costs by 70%; API priced at $0.15 input / $0.60 output per 1M tokens
- Self-reported benchmarks: 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, and 82% on CyberGym
- Live on API and MiniMax Code; open model weights scheduled for public release on October 23
- Validation imperative: benchmark metrics remain internal disclosures requiring independent audit via LMArena and Artificial Analysis
“Production agent efficiency is not governed solely by raw token pricing, but by the minimization of failure-recovery retry loops required to complete a complex engineering objective deterministically.”
The release of open weights later this week represents a major catalyst for independent developers seeking to run high-performance, self-hosted coding agents with unprecedented unit economics.
References & Primary Sources
Related Analysis & Context
Meta Muse Spark Open: 816B Model on Apple M5 Ultra
Meta has launched the open weights for Muse Spark Open, featuring an 816B MoE parameter footprint, a 1M token context window, and native local execution on desktop Apple M5 Ultra Mac Studios using the MLX framework.
FLUX 3 Image Klein: 5.6B Open-Weight 4K Visual Model
Black Forest Labs has released FLUX 3 Image Klein, a 5.6B parameter open-weight visual generation and editing model featuring native 4K synthesis, 10-image reference conditioning, bounding-box control, and local execution on 8GB VRAM GPUs.
Inside Meta Prometheus: The Gigawatt-Scale Supercluster Scaling to 1M GPUs
Engineering at Meta has published the technical blueprint for its gigawatt-scale Prometheus supercluster in New Albany, Ohio, engineered to interconnect up to 1 million NVIDIA GB200/GB300 GPUs with 51.2T RoCE fabrics, >85% training goodput, and Dynamo Dimmer grid stabilization.