MiniMax M3.1 Flash: API Setup, Thinking and Costs

MiniMax announced M3.1 Flash on October 19, 2026, with access through its API, M Plan and MiniMax Code. An application subscription and a pay-as-you-go API account have different usage arrangements; use the API rates below for calls billed through the Open Platform.

MiniMax identifies Space Bunny (SpaceBunny) as an early version of M3.1 Flash. The exact MiniMaxAI/MiniMax-M3.1-Flash weights are now downloadable. The model card lists approximately 428B total parameters and 23B activated per token, with text, image and video inputs. Use the official serving recipes for local inference; installing the coding client alone still uses a hosted model.

The release combines MiniMax Sparse Attention throughout the model, quantization-aware training and an included DSpark draft head. The model card links deployment recipes for SGLang, vLLM and TokenSpeed. Start with the runtime version and flags for your GPU architecture and checkpoint format.

For a capacity sanity check, 428B parameters at two bytes each are about 856 GB of weight data; at one byte, about 428 GB; at half a byte, about 214 GB. These are decimal weight-only arithmetic, before mixed-precision tensors, scales, the draft head, KV cache, vision processing and serving buffers. The 23B active count describes per-token computation and cannot replace the total in a memory budget.

The vLLM guide supplies different recipes for TP4 GB200, TP4 MI355X and TP8 H100/H200. Its Hopper example limits context to 262,144 tokens, disables video input and does not enable DSpark. A model-level 1M context and video capability therefore do not transfer unchanged to every server recipe.

In an October 23 Hugging Face report, mmssix describes four RTX PRO 6000 Blackwell GPUs with the vllm/vllm-openai:minimax-m31-flash image, vLLM 0.1.dev23341+g8744dd2fb and checkpoint revision 54e4e69. Non-streaming split reasoning correctly, while streaming included thinking markup in content. The same report needed FP8 cache settings and an explicit draft attention backend on SM120. Treat this as a version-specific integration report; test streaming and tool parsing on the selected build before exposing them to users.

Commercial use requires visible attribution and a one-time notice to MiniMax. Products or services exceeding $20 million in annual revenue require prior written authorization. The license also contains prohibited-use conditions. Read the exact checkpoint license before choosing self-hosting for a commercial service; downloadable weights do not make this an Apache or MIT release.

MiniMax’s direct API rates, checked October 20, are in USD per million tokens. Cached input means eligible cache reads, rather than every repeated word in a conversation. Use the provider’s usage accounting to separate cache hits from fresh input.

At $0.15 for fresh input, $0.015 for cached input and $0.60 for output, an assumed run with 100,000 fresh input tokens, 900,000 cached tokens and 20,000 output tokens costs $0.0405. Without those cache hits, the same one million input tokens and 20,000 output tokens cost $0.162. These are token-budget examples, not measured task bills.

Add together every model call in an agent run. Repository context and tool results can be sent again on later turns, while retries produce additional usage. Reconcile thinking and answer generation with the billable output reported by the provider rather than counting only the final answer shown to the user. Tool services and your own execution environment have separate costs.

Create a MiniMax Open Platform API key and keep it on your server. Install the OpenAI Python package, set MINIMAX_API_KEY in your environment and use the MiniMax base URL. This example asks for a review plan; it does not grant a model access to your repository or execute tools.

Check the returned answer against the requested files and tests before connecting a tool loop. For multi-turn tool use, preserve the complete assistant message, including tool calls and reasoning_content, as described in MiniMax’s integration guide. Dropping these fields can break reasoning continuity.

M3.1 Flash supports low, medium, high, xhigh and max effort; omission selects max. Thinking cannot be disabled: reasoning_effort="none" or a disabled thinking setting returns an error. An OpenAI-compatible endpoint therefore does not mean every setting from another provider transfers unchanged.

Start with one effort setting and a small set of tasks with known acceptance criteria. Record successful completions, generated tokens and elapsed time. Then try a lower effort on the same inputs. A shorter response is useful only if it still produces the correct patch, document or tool action.

MiniMax reports more than 200 output tokens per second in some everyday-use conditions and a median above 170, alongside coding and agent benchmarks. These are publisher results, not a latency estimate for your connection or a completed coding task. A long prompt, reasoning effort and external tools can dominate the time a user waits.

Artificial Analysis’s October 24 model-page snapshot lists 87.2 output tokens per second for MiniMax’s API and flags high output-token use in its Intelligence Index evaluation. Its benchmark task cost accounts for input, cache, reasoning and answer usage. This is a separate test from MiniMax’s launch measurement; use the index to investigate token use and task cost rather than treating either generation-rate figure as an end-to-end agent time.

For an application comparison, keep the task, tool permissions, retry cap and completion criteria fixed. Compare total spend per accepted result with the existing DeepSeek workflow or another hosted text model, instead of ranking providers by isolated generation speed.