http://x.com/i/article/2078183632641089536

Running MiniMax M3 in Production with Together AI Provisioned Throughput

A practical guide to reserved token capacity, PTU sizing, deployment options, and workload economics.

Last week, Together AI introduced Provisioned Throughput for MiniMax M3. The offering gives teams a way to reserve M3 inference capacity in tokens per minute, backed by a 99% uptime SLA, while Together AI manages the underlying serving infrastructure.

The launch adds a new deployment option for teams moving M3 workloads into production. Serverless remains available for usage-based access, while Dedicated Inference is designed for custom or fine-tuned models that require deeper control over the serving environment. Provisioned Throughput sits between them: managed inference for a standard model, with capacity reserved in advance.

This guide explains where Provisioned Throughput fits, how capacity is measured for M3, and how workload shape affects the economics.

Where Provisioned Throughput fits

Together AI offers three inference deployment models: Serverless Inference, Provisioned Throughput, and Dedicated Inference.

Serverless Inference

- Capacity model: Pay for the tokens used, with best-effort capacity

- Best suited for: Starting development and workloads that do not require reserved capacity

Provisioned Throughput

- Capacity model: Pay for reserved tokens-per-minute capacity, with a 99% uptime SLA

- Best suited for: Production workloads on standard models that require guaranteed capacity

Dedicated Inference

- Capacity model: Reserve dedicated infrastructure

- Best suited for: Fine-tuned or custom models requiring deeper control over the serving environment

The inference API on Together AI is identical across Serverless Inference, Provisioned Throughput, and Dedicated Inference. This allows teams to use the same API surface while selecting the capacity model that fits their workload.

How Provisioned Throughput works

Provisioned Throughput is purchased in Provisioned Throughput Units, or PTUs.

Each PTU is a fixed amount of model capacity held exclusively for the customer. One PTU costs $0.05 per minute and provides a guaranteed rate of tokens per minute for the selected model. Teams purchase capacity rather than individual GPU hours, while Together AI operates the infrastructure behind the service.

Inference traffic is divided into three categories:

- uncached input tokens;

- cached input tokens;

- output tokens.

Each category consumes PTU capacity at a different rate. For MiniMax M3, one PTU provides:

A workload can use any combination of the three. Cache-efficient workloads can process more total tokens within the same reserved capacity, while output-heavy workloads consume capacity more quickly. The workload mix changes the rate at which PTUs are consumed, but not the uptime SLA.

Using Together AI’s published conversion rates, a mixed workload can be estimated as:

PTUs required ≈ uncached input TPM ÷ 138,840

- cached input TPM ÷ 694,200

- output TPM ÷ 23,140

The result should be rounded up to the next whole PTU when sizing reserved capacity.

Economics at full utilization

At full utilization, MiniMax M3 on Provisioned Throughput costs approximately:

Effective M3 Rate by Traffic Type

- Input: $0.36 per million tokens

- Output: $2.16 per million tokens

By comparison, Claude Opus 4.8’s standard list prices are $5 per million input tokens and $25 per million output tokens.

Because PTUs reserve capacity in advance rather than charging strictly by usage, these figures represent effective rates at full utilization. A team’s realized cost per token will depend on how consistently it uses its committed capacity and on the mix of input, cached input, and output traffic.

Economics across representative workloads

Together AI modeled three production workloads to show how request volume, token mix, and cache reuse affect PTU requirements and effective cost.

🛠️ Workload 1: Coding Agent

- Traffic Assumptions: 85K input & 450 output tokens per request; 80% cache hit; 1 request per second

- M3 on Provisioned Throughput: $33K / month

- Claude Opus 4.8 (List Price): $342K / month

- Estimated Savings: ~90% lower

📞 Workload 2: Support Agent

- Traffic Assumptions: 3K input & 500 output tokens per request; 50% cache hit; 5 requests per second

- M3 on Provisioned Throughput: $24K / month

- Claude Opus 4.8 (List Price): $273K / month

- Estimated Savings: ~91% lower

📄 Workload 3: Document Pipeline

- Traffic Assumptions: 15K input & 800 output tokens per request; 20% cache hit; 2 requests per second

- M3 on Provisioned Throughput: $35K / month

- Claude Opus 4.8 (List Price): $428K / month

- Estimated Savings: ~92% lower

These modeled estimates assume continuous provisioning, PTUs priced at $0.05 per minute, and Claude Opus 4.8 list pricing for standard input, cache reads, and output. Actual costs will vary based on utilization, traffic shape, and cache behavior.

PTU requirements depend on workload shape, not request volume alone. The coding-agent profile benefits from high cache reuse, while the support-agent profile places more weight on output capacity and the document pipeline is driven more heavily by uncached input. Teams should therefore size Provisioned Throughput using their actual mix of uncached input, cached input, and output tokens, together with request rate and cache behavior to improve utilization and avoid overprovisioning.

Why this matters for M3 workloads

MiniMax M3 was designed for coding and agentic work, supports context windows of up to one million tokens, and natively accepts text, image, and video input. Its MiniMax Sparse Attention architecture was developed to make long-context processing more computationally efficient.

These capabilities support workloads whose inference patterns can change over the course of a session. A coding agent may repeatedly reuse repository context while producing tool calls, code edits, test output, and reasoning. A document workflow may process substantial new input on every request. Other applications may generate comparatively more output.

Provisioned Throughput gives teams a way to reserve M3 capacity around those workload patterns, rather than planning directly around GPU configurations.

As Linda Sheng, President of Global Business at MiniMax, shared:

“Our models are built for demanding production workloads and Together AI’s Provisioned Throughput gives teams a way to run them with the reliability those workloads require. As teams shift volume off closed APIs, Together AI’s infrastructure lets them adopt M3 at scale with commitments they can build a business on. That’s exactly the kind of access we want for our models.”

Available now

Together AI Provisioned Throughput is available for MiniMax M3, with capacity offered across North America, EMEA, and additional regions. The minimum term is one month, and Together AI states that discounts are available at higher commitment levels.

Teams can use Together AI’s calculator to estimate PTU requirements using their own request volume and mix of input, cached input, and output tokens.

Ready to run MiniMax M3 with predictable production capacity?

Estimate your PTU requirements
Read the Provisioned Throughput documentation
Talk to Together AI