MiniMax M3.1 Flash Lands on Ollama Cloud at One-Third Prior Cost with 1M Context

MiniMax M3.1 Flash is a frontier-level AI model from MiniMax that targets agentic tasks coding and visual reasoning with a 1 million token context window and image support at reduced cost.

MiniMax officially rolled out MiniMax M3.1 Flash on October 19 2026 through its API service the M Plan subscription and the MiniMax Code platform. The release supplies nearly three times the inference speed of the predecessor M3 model together with a seventy percent reduction in usage cost for agentic workloads. Developers obtain immediate access to the new capabilities through existing accounts while new users register via the listed channels.

The model maintains flat API pricing across all usage volumes. Input tokens cost 0.15 dollars per million cached input tokens cost 0.015 dollars per million and output tokens cost 0.60 dollars per million. These rates apply uniformly without tiered surcharges according to the company pricing schedule.

What benchmark scores place MiniMax M3.1 Flash at frontier level?

MiniMax reports 73.7 percent accuracy on the DeepSWE v1.1 benchmark for the new model. Additional task scores include 62 percent on Automation-Bench within the Cowork category 30 percent on ALE 74 percent on the DeepSWE coding evaluation and 91 percent on BabyVision for visual reasoning. These figures reflect performance across automation coding and visual domains at one third the cost of the prior generation.

Output throughput reaches more than 200 tokens per second with a P50 measurement above 170 tokens per second during everyday operation. The combination of benchmark results and measured speed supports deployment in production agent systems where both accuracy and latency matter.

Which technical specifications define the model architecture?

MiniMax M3.1 Flash provides a 1 million token context window that accommodates entire code repositories or lengthy documents within a single prompt. Image understanding enables direct processing of visual inputs alongside text. Tool calling and reasoning functions integrate with external systems while a tunable thinking depth parameter allows users to adjust computational effort per query.

Key specifications and pricing for MiniMax M3.1 Flash

Metric	Value	Source Attribution
Input price per million tokens	$0.15	MiniMax pricing announcement
Output price per million tokens	$0.60	MiniMax pricing announcement
Cached input price per million tokens	$0.015	MiniMax pricing announcement
Context window	1 million tokens	MiniMax and Ollama listings
Output speed	200+ tokens per second P50 170+	MiniMax performance data
Benchmark DeepSWE v1.1	73.7 percent	MiniMax announcement

How does Ollama Cloud hosting operate for this model?

Ollama lists the hosted variant as minimax-m3.1-flash:cloud with approximately 976 thousand tokens of effective context. The service supplies vision tool use and thinking capabilities through a US data center. All requests follow a strict zero data retention policy that prevents storage of prompts or completions.

MiniMax M3.1 Flash is on Ollama! 🔒 Hosted in the US and with zero data retention. Fast coding and agentic workflows, with image understanding and a 1M-token context window.Ollama, Official account

What is the planned release timeline for weights and further access?

1. October 19 2026 marks the official rollout through the MiniMax API M Plan and MiniMax Code.
2. October 23 2026 is the scheduled date for public release of the model weights.
3. Immediate availability exists on Ollama Cloud for hosted inference without local deployment.
4. Post release independent evaluations are expected to verify reported benchmark figures.

What market and stakeholder implications arise from the pricing model?

The two thirds cost reduction relative to M3 lowers barriers for organizations that run high volume agentic or coding workloads. Cached input pricing at 0.015 dollars per million further reduces expense for repeated context retrieval. Enterprises gain a predictable cost structure that scales linearly with token consumption.

Open weights availability on October 23 enables local deployment fine tuning and academic study. Research teams obtain the ability to inspect architecture details and conduct custom evaluations without ongoing API fees. Platform operators may integrate the weights into additional hosting environments once the files circulate.

What expert and community reactions have followed the announcement?

The official Ollama account highlighted the hosted option and zero retention policy in a public post. The MiniMax company account emphasized frontier level scores across automation coding and visual tasks at one third the prior cost and three times the speed. Both statements focus on concrete access and performance details rather than qualitative claims.

What developments are expected after the October 23 weights release?

Independent verification of the 73.7 percent DeepSWE score and other benchmarks will likely appear once weights reach public repositories. Additional platform integrations beyond Ollama Cloud may follow as developers experiment with local inference. Refinements to the tunable thinking depth parameter could occur based on usage patterns observed through the API.