Kimi K3 vs MiniMax M3: Price, Output Length and Media

All figures below are from the China AI Hub database, last verified 2026-09-22. Where a field is not publicly disclosed, we say so rather than estimating.

At a glance

Dimension	Kimi K3	MiniMax M3
Context window	1,048,576 tokens	1,048,576 tokens
Maximum output	1,048,576 tokens	Not publicly disclosed
Input price (per 1M)	$3.00	$0.30
Output price (per 1M)	$15.00	$1.20
Open weights	Yes (Kimi K3 License)	Yes (MiniMax Community License)
API available	Yes	Yes
Reasoning	Yes	Yes
Coding	Yes	Yes
Vision / Video	Yes / Yes	Yes / Yes
Tool calling	Yes	Yes
Agent capability	Yes	Yes
Benchmark records in DB	6	5

Pricing

The widest price spread among our current flagship comparisons: Kimi K3 lists $3.00 input / $15.00 output per 1M tokens; MiniMax M3 lists $0.30 input / $1.20 output — one-tenth of Kimi K3’s input price and one-twelfth of its output price. Both are official pay-as-you-go API prices. On long reasoning or generation workloads the per-task cost difference can be an order of magnitude.

Context and output

Both list a 1,048,576-token context window. Kimi K3 lists a 1,048,576-token maximum output (the full window); MiniMax M3 does not publicly disclose a maximum output figure. For very long generations the listed difference is material.

Capabilities

Capability profiles are close: both list reasoning, coding, vision, video, tool calling and agent capability. Neither lists structured output or function calling under those exact labels. Media coverage (vision and video) is listed for both.

Openness and licensing

Both are open weight, with different obligations. Kimi K3 uses the Kimi K3 License: MIT-style terms, plus a separate agreement for Model-as-a-Service operators above $20M revenue and a UI attribution requirement above 100M MAU / $20M monthly revenue. MiniMax M3 uses the MiniMax Community License: free for non-commercial use; commercial use requires “Built with MiniMax M3” attribution and, above $20M yearly revenue, written authorization (otherwise a one-time notice to MiniMax). Read both texts before commercial self-hosting — the non-commercial restriction in the MiniMax license is the sharper boundary.

Benchmarks and verification

The database holds 6 benchmark records for Kimi K3 and 5 for MiniMax M3, labeled by source type and version. Different benchmarks cover different models; we do not rank the two on this page. Check each benchmark page for score-by-score source labels.

Benchmark comparability is limited: the two models’ scores come from different test configurations — different benchmark versions, evaluation harnesses and tool setups — and several benchmarks cover only one of the two. Do not rank on headline scores alone.

Why each difference matters

China AI Hub analysis indicates the following per-dimension implications, drawn from the listed facts above.

• Cost: MiniMax M3 has an advantage at ~1/10 the input and ~1/12 the output price; on long reasoning or generation the per-task cost difference can be an order of magnitude.
• Long-context: both list a 1M-token input window; Kimi K3 has an advantage in very long generations because it lists the full-window 1M-token maximum output, while MiniMax M3 has a limitation in not publicly disclosing one.
• Reasoning: both list reasoning, so neither has a listed advantage — cost and output length, not capability, differentiate on this axis.
• Coding: both list coding, and the capability profiles are closely matched on paper.
• Agent / media: both list agent capability, vision and video, so this axis does not differentiate.
• Deployment / license: both are open-weight but with different obligations; MiniMax M3 has a limitation in its non-commercial boundary, and Kimi K3 in its revenue-triggered obligations — choose by which license you can actually operate under.
• API: both expose an API; cost and output-length profile decide.

Trade-off summary

• Price: MiniMax M3 lists roughly one-tenth the input and one-twelfth the output price of Kimi K3.
• Output length: Kimi K3 lists a 1M-token maximum output; MiniMax M3 does not disclose one.
• Capabilities: closely matched on paper — reasoning, coding, vision, video, tools, agent.
• License: both custom; MiniMax’s has a non-commercial boundary, Kimi’s has revenue-triggered obligations.

Choose by cost profile and output-length needs, and by which license you can actually operate under. Verify current prices on the official pages before committing.

Decision context

China AI Hub analysis indicates the following decision-context implications, drawn from the listed facts above.

For API developers. MiniMax M3 lists ~1/10 the input and ~1/12 the output price — on long reasoning or generation the per-task cost difference can be an order of magnitude. Capability profiles are closely matched (both list reasoning, coding, vision, video, tool calling and agent capability), so cost and output-length profile decide. Kimi K3 lists a full 1M-token maximum output where MiniMax M3 does not disclose one. Latency is not publicly documented on this page.

For self-hosting. Both are open weight with custom licenses — the Kimi K3 License (revenue-triggered obligations) versus the MiniMax Community License (non-commercial boundary). Choose by which license you can actually operate under. Hardware requirements and quantization are Not publicly documented on this page.

For coding agents. Both list coding, tool calling and agent capability, so this axis does not differentiate. Both list a 1M-token context. SWE-benchmark and terminal scores are not broken out (aggregate counts 6 vs 5).

For enterprise. Region, SLA, data-residency and compliance terms are not publicly documented — confirm with the vendor. Deployment is comparable (both self-hostable open weight), but the licenses differ: MiniMax’s has a non-commercial boundary, Kimi’s has revenue-triggered obligations — verify which you can legally operate under.

What is uncertain

• MiniMax M3 does not publicly disclose a maximum output figure.
• Benchmark comparability is limited — different versions, harnesses and partial coverage.

Sources

• Moonshot AI — Kimi K3 (GitHub)
• Kimi platform — Models documentation
• MiniMax — Models introduction
• MiniMax — MiniMax M3 announcement

Labels used above: Official fact (prices, context windows, capabilities and license terms from primary provider sources), Vendor-reported claim (benchmark scores), and China AI Hub analysis (the “why each difference matters” reasoning, introduced as analysis). No third-party benchmark evidence is currently recorded for these models.