MiniMax M3

MiniMax M3 is a multimodal MoE model with a context window of one million tokens, focused on agentic workflows, coding, and tool use. Of 428 billion total parameters, only 23 billion are active per token; the model processes text, image, and video as input. Its Chinese origin requires a separate data privacy risk assessment when used via cloud.

MiniMax Version m3 Commercial use permitted MoE 428 B (23 B active) 1000 K Context 05/2026 $0.3 / $1.2 per 1M

Open Weights
Frontier
OR
Text
Vision
Video
Interactive

Sovereign Risk: HIGH MiniMax is a Chinese company and subject to China’s National Security Law (NSL), which may enable state access to data. The model has been released as open weights, but remains high-risk from a sovereignty perspective when data or workflows are processed under Chinese jurisdiction.

Tool-use details

Asset performance, reliability, and runtime profile

Name	P1	P2	Combined
EU License Research	100	80	88
HTTP Fetch & Extract	80	80	80
Tool Failure Handling (404)	80	80	80
Web Search & Tool Selection	100	80	91
URL Construction & Fetch	80	35	57.5
Multilingual Search & Synthesis	100	20	60

Name

Combined

EU License Research

100

HTTP Fetch & Extract

Tool Failure Handling (404)

Web Search & Tool Selection

100

URL Construction & Fetch

57.5

Multilingual Search & Synthesis

100

CrucibleMark evaluates tool use across 6 independent tests. Click a test name for details.

Reliability

Tool Call Valid: No

Retry: Not required

Hallucination: Detected

Reliability measures how consistently a model actually executes tool calls: Tool Call Valid schema and format accepted, Retry Required successful only after retry, Hallucination Flag fabricated tools or parameters detected. All three green means production-ready.

Runtime profile

Call 1: 5.28; First Request

MCP: 0.82; Protocol Latency

Synthesis: 20.54; Response Generation

Total: 159.85; Sum of All Phases

Token: 10037; Input + Output

Cost: $0.0063; Cost per Run

The runtime profile shows the latency and cost metrics for the model run: Call 1 First Request, MCP Protocol Latency, Synthesis Response Generation, Total sum of all phases, supplemented by Token input and output and Cost cost per run.

Deployment Verdict

Conditional deploy, because tool execution is strong and protocol-compliant, but the detected hallucination at only moderate synthesis fidelity limits confidence in content-sensitive tool pipelines.

Tool Execution Profile

MiniMax M3 behaves in a production-near manner at the execution level. Tool calls are valid, MCP-compliant, and completed without retry. This argues against a formatting issue and in favor of stable tooling behavior. The model is particularly strong where it must identify the right tool on its own: in the Web Search & Tool Selection test, which checks whether search is chosen over fetch without any hint, it makes the correct decision reliably. This demonstrates genuine tool intelligence rather than a rigid call schema.

It is less reliable on the URL Construction test, which requires deriving the target URL from internal knowledge and then fetching it correctly. Here it performs adequately, but not precisely enough for deterministic pipelines with strict URL requirements. The pattern is clear: when the problem lies in choosing the right tool, M3 is strong. When it must accurately construct its own facts before the tool call, reliability drops.

Synthesis Fidelity

How well does it condense tool results? Only adequately. The P2 performance shows that MiniMax M3 often correctly aggregates retrieved content, but does not consistently prioritize or condense it cleanly. This is most visible in the Multilingual Search & Synthesis test, which requires cross-lingual research with German-language output: the research succeeds, but the final synthesis breaks down noticeably. For pipelines where the model is expected to aggregate rather than merely cite, this is an operational weakness.

Does it stay within tool results or fall back on training data? In the Honeypot EU License Research test, which checks whether current license restrictions are actually retrieved from web sources, it stays within the verified tool result. Content Verification State A and no hallucination are a strong trust signal. At the same time, the overall hallucination finding remains a security risk: once a model outputs fabricated facts as a result within a tool chain, it is not just a single answer that degrades — the reliability of the entire infrastructure is undermined.

Error Resilience

In the Tool Failure Handling (404) test, which measures transparent behavior on a failed retrieval, MiniMax M3 responds acceptably. It does not hallucinate page content despite the error and communicates the failure in a comprehensible way. This is critical for production, because the pipeline can then handle the error correctly downstream.

Operational Profile

5.28s first call, 20.54s second call, 159.85s total. Slow for the performance shown. MCP latency 0.82s is non-critical. Cost per run 0.006320 USD. Inexpensive for a Frontier agent model.

Conclusion & Recommendation

Suitable for agentic pipelines with clear tool boundaries, web research, error propagation, and controlled extraction. Not suitable for compliance-adjacent or multilingual synthesis pipelines where every condensed statement must be reliably traceable to tool results. Deploy when orchestration is strong and a downstream verification step exists. Do not use as an unsupervised synthesis endpoint.

This analysis was generated automatically based on benchmark data. Model used: GPT 4.5 by OpenAI. The raw data and full methodology are documented in the GitHub project.