MiniMax M3.1 Flash Released: A 428B MoE Agent Model with Input at Just $0.15/M

MiniMax has launched M3.1 Flash, a 428B-parameter MoE (Mixture of Experts) model with roughly 23B active parameters and a context window of 1 million tokens. API pricing is set at $0.15 per million input tokens and $0.60 per million output tokens, with weights planned for release on October 23.

1. The "Parameter Magic" of Sparse Architecture

The most interesting aspect of M3.1 Flash is its parameter structure: 428B total parameters, but only about 23B activated per inference. This means model capacity and inference cost are deliberately decoupled—it retains the knowledge density that comes with large-scale parameters while keeping actual computation close to the level of a mid-sized dense model.

MoE is not a new concept, but keeping the activation ratio at around 5% (23B/428B) is a relatively aggressive design among current open and semi-open models. This trade-off typically serves one goal: driving down the compute cost per token in long-context and high-concurrency scenarios without sacrificing too much complex reasoning capability.

MoE and Activation Ratio

MoE (Mixture of Experts) replaces the single feed-forward network (FFN) in a traditional Transformer with multiple parallel "expert" subnetworks, and adds a router network that computes weights for each expert per token, allowing only the few highest-scoring experts to participate in computation and then combining their outputs with weights. As a result, the model's total parameters determine how much knowledge it can store, while active parameters depend only on the number of experts actually selected each time—the two can be decoupled.

The industry commonly uses the "activation ratio" (active parameters/total parameters) to describe this sparsity. Dense models always have an activation ratio of 100%; mainstream MoE models typically fall in the 10%–20% range—for example, designs with hundreds of billions of total parameters and tens of billions of active parameters. M3.1 Flash pushes the activation ratio down to about 5%, meaning the floating-point operations per token are significantly lower than a dense model with the same total parameters—this is precisely the structural prerequisite that enables its low API pricing. The costs include routing overhead itself, imbalanced expert load (some experts being over-invoked), and VRAM usage still calculated based on total parameters.

2. 1M-Token Context and the "Agentic" Label

Officially, M3.1 Flash is positioned as an "Agentic" model with a 1M-token context. Taken together, these two features point to the same class of workloads: agent tasks that require maintaining state over long periods, frequently calling tools, and accumulating memory across multi-turn interactions.

The significance of long context for agents goes beyond "fitting more text." When a model needs to read a codebase, browse multiple documents, or maintain a task plan within a single session, the actual usable context length directly determines how complex a task it can complete. A 1M-token window allows a single session to cover project materials of considerable scale, reducing the engineering burden of external retrieval and state management.

However, there is often a gap between context length and actual effective utilization. Attention decay in ultra-long contexts and the ability to locate key information are still aspects that require empirical verification—the officially announced window size cannot be directly equated with "long-context reasoning quality."

The Relationship Between Agentic Models and Long Context

"Agentic" is not a strict academic definition; it refers to models being used in agent workflows: the model doesn't just answer single-turn questions, but autonomously decomposes goals, calls external tools such as search/code execution/file reading and writing, adjusts plans based on tool return results, and continuously advances tasks through multiple rounds of iteration.

This type of workload has three special requirements for context: first, state persistence—task plans, completed steps, and intermediate results must remain in context, otherwise the model will repeat work or lose its goal; second, material scale—reading codebases and reviewing multiple documents often brings in hundreds of thousands of tokens at once; third, tool output expansion—the return value of each tool call is appended to context, rapidly consuming the window. Therefore, the value of long context windows for agents is more about "reducing external memory management engineering" rather than simply accommodating long text. For this reason, the gap between the nominal window size and effective utilization is amplified in agent scenarios—forgetting a key early instruction can cause an entire task chain to go off track.

3. Pricing Strategy: What $0.15 Input Means

At $0.15 per million input tokens and $0.60 per million output tokens, this price point falls clearly in the low-price range among 400B-class models. Combined with the sparse activation architecture, it can be inferred that MiniMax made targeted optimizations to inference costs and passed the savings on to API pricing.

For developers, the 1:4 price ratio between input and output is also worth noting. Agent scenarios often require repeatedly reading large amounts of context (input-intensive) while producing relatively brief output—this pricing structure is more favorable to the input side, making it suitable for read-heavy tasks such as retrieval augmentation, codebase analysis, and document batch processing.

Why Input/Output Token Pricing Differs

LLM APIs generally charge higher unit prices for output tokens (commonly 3–5 times that of input), because the computational characteristics of the two differ. The input side is the prefill phase: all input tokens can be computed in parallel, GPU utilization is high, and identical prefixes can be reused through prefix caching, resulting in low marginal cost. The output side is the decode phase: token-by-token autoregressive generation, where each generated token requires a full forward pass and reading of all weights and KV cache—limited by memory bandwidth, this is a typical memory-bound workload with far lower throughput than prefill.

Therefore, the 1:4 input-output price ratio is an industry norm rather than an exception. For read-heavy applications such as agents, RAG, and code analysis, costs are mainly determined by the input side, so the actual benefit of low-cost input is amplified; conversely, bills for long-form generation applications will fall more on the output side. When evaluating cost-effectiveness, one should estimate the blended unit price based on the input-output ratio of one's own business, rather than looking only at the input quote.

4. Weight Release Timing and Reasons to Wait

Officially, weights will be released on October 23. Before the weights actually land, judgments about M3.1 Flash should remain open:

• The model card and license terms have not yet been published; commercial restrictions and distribution conditions are unclear
• Third-party benchmarks are lacking; the actual capability ceiling of 428B total parameters remains to be verified
• The recall quality and inference latency of the 1M context in real tasks require independent evaluation

Based on the release information itself, the core selling points of M3.1 Flash are the combination of four factors: "large capacity, low activation, low price, long context." Whether this combination translates into actual agent productivity depends on community empirical testing after the weights are released, not on numbers in a parameter table.

For teams evaluating agent infrastructure, this is an option worth adding to the comparison list, but before benchmarks and license information are complete, it is premature to treat it as a confirmed "king of cost-effectiveness."

FAQ

What are the active parameters of M3.1 Flash? About 23B, with 428B total parameters, using a sparse MoE architecture.

How is API pricing calculated? $0.15 per million input tokens and $0.60 per million output tokens.

How large is the context window? Officially, 1 million tokens.

When will the weights be released? Planned for October 23, with the specific license not yet announced.

One-line Summary

MiniMax releases the 428B-parameter MoE model M3.1 Flash, with 23B active parameters, 1M context, and extremely low pricing.

Paragraph Summary

MiniMax has launched M3.1 Flash, a sparse MoE model with 428B total parameters and about 23B active parameters—an activation ratio of roughly 5%—and a context window of 1 million tokens, officially positioned as an Agentic model. API pricing is $0.15 per million input tokens and $0.60 per million output tokens, clearly in the low-price range among 400B-class models; the 1:4 input-output price ratio is more favorable for read-heavy tasks. Weights are planned for release on October 23. The article notes that the value of long context for agents lies in reducing external state management burden, but a nominal window does not equal actual long-context reasoning quality; until the model card, license, and third-party benchmarks are published, its cost-effectiveness still requires community empirical verification.