MiniMax M3 vs M2.7: Same Price, Twice the Context

All figures below are from the China AI Hub database, last verified 2026-09-20 (M2.7 last verified 2026-09-27). Where a field is not publicly disclosed, we say so rather than estimating.

At a glance

Dimension	MiniMax-M3	MiniMax-M2.7
Status (database)	active	active
Context window	1,048,576 tokens	204,800 tokens
Maximum output	Not publicly disclosed	Not publicly disclosed
Input price (per 1M)	$0.30	$0.30
Output price (per 1M)	$1.20	$1.20
Open weights	Yes (MiniMax Community License)	Yes (custom non-commercial)
API available	Yes	Yes
Reasoning	Yes	Yes
Coding	Yes	Not listed
Vision / Video	Yes / Yes	No / No
Tool calling	Yes	Yes
Agent capability	Yes	Not listed
Benchmark records in DB	5	2

Pricing

Both list the identical price — $0.30 input / $1.20 output per 1M tokens — with cache reads at $0.06. MiniMax-M3’s pricing splits at the 512K input boundary (double above), and it offers a priority tier at 1.5x with a permanent 50%-off list. On price the two are equivalent per the database, which makes the capability differences the deciding factors.

Context and output

MiniMax-M3 lists a 1,048,576-token context window (with at least 512K guaranteed usable and pricing split at 512K). MiniMax-M2.7 lists a 204,800-token context. Neither publicly discloses a standalone maximum-output ceiling (M3’s 131,072 is derived from the official card’s evaluation config). For long-context work the difference is a 5x listed input window.

Architecture and parameters

MiniMax-M3 is a ~428B-total / ~23B-activated MoE built around MiniMax Sparse Attention (MSA), which MiniMax claims delivers 9x prefill and 15x decode speedup versus M2 at 1M context — a vendor claim, not an independent measurement. MiniMax-M2.7 does not disclose its parameter count; it is positioned as MiniMax’s “self-evolving” model and, unlike M3, is not documented with a sparse-attention architecture card. The architecture disclosure gradient between the two is itself a recorded difference.

Coding

MiniMax-M3 lists coding and publishes vendor-reported SWE-bench Pro (59.0) and Terminal-Bench 2.1 (66.0). MiniMax-M2.7 does not list coding in its recorded capability set (reasoning and tool calling only). For coding workloads the database records a clear difference in listed capability.

Vision

MiniMax-M3 lists vision and video input; MiniMax-M2.7 is text-only input. For multimodal understanding only M3 lists the capability.

Agent capabilities

MiniMax-M3 lists agent capability and tool calling, and publishes a vendor-reported MCP Atlas score (74.2) measuring MCP-connected tool use — the pattern agent builders rely on. MiniMax-M2.7 lists tool calling but not agent capability. Thinking is disabled by default on M3 (thinking=adaptive enables it); on M2.7 interleaved thinking is always on and cannot be disabled via API.

License and openness

Both are open weight, but the licenses differ materially. 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). MiniMax-M2.7 uses a custom NON-COMMERCIAL license: any commercial use requires prior written authorization from MiniMax, with attribution required. The M2.7 license is the stricter boundary — commercial self-hosting is effectively gated for both, but M3 allows commercial use below the revenue threshold with attribution.

Deployment

Both are self-hostable open weight and served through the MiniMax Platform. Both list regions china and international. M3 is a ~428B-total / ~23B-active MoE; M2.7’s parameter count is not publicly disclosed. Self-hosting hardware requirements and quantization are not publicly documented for either.

Benchmarks

The database holds 5 benchmark records for MiniMax-M3 (BrowseComp, PostTrainBench, SWE-bench Pro, Terminal-Bench 2.1, MCP Atlas) and 2 for MiniMax-M2.7 (GDPval-AA ELO, MM Claw end-to-end). The sets do not overlap. All are vendor-reported.

Benchmark comparability is limited: the two models’ scores come from different benchmark sets and configurations, and the M2.7 rows (an ELO rating and a custom end-to-end benchmark) are not directly comparable to M3’s accuracy scores. 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.

• Pricing: identical listed prices ($0.30/$1.20), so this axis does not differentiate — capability fit decides.
• Context: MiniMax-M3 has an advantage in long-context work at 1,048,576 tokens versus M2.7’s 204,800.
• Coding: MiniMax-M3 is more relevant for coding workloads — it lists coding and coding benchmarks where MiniMax-M2.7 lists none.
• Vision / media: MiniMax-M3 is more relevant when image or video input is required — it lists both where M2.7 is text-only.
• Agent: MiniMax-M3 is more relevant for agent/tool workloads — it lists agent capability and MCP tool use where M2.7 lists only tool calling.
• License: both conditional-open; M2.7 has a limitation in its strictly non-commercial boundary, while M3 allows commercial use with attribution below the revenue threshold.
• Deployment: both self-hostable; the footprint differs (M3 discloses ~428B/~23B; M2.7 does not disclose parameters).
• Benchmark: M3 records 5 and M2.7 records 2 vendor-reported rows with no overlap; not directly comparable.

Trade-off summary

• Context: MiniMax-M3 lists a 1M-token window; M2.7 lists 204,800.
• Capabilities: MiniMax-M3 lists coding, vision, video and agent capability; M2.7 lists reasoning and tool calling only.
• Price: identical ($0.30/$1.20).
• License: M3 community license (commercial with attribution below threshold) versus M2.7 non-commercial-only.

Choose by workload: long-context, multimodal, coding and agent workloads favor MiniMax-M3’s listed capabilities; M2.7’s niche is its self-evolving positioning, and it is priced the same but with a stricter license. 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. The prices are identical, so the choice is capability-driven: MiniMax-M3 lists coding, vision/video and agent capability plus a 1M context; M2.7 lists only reasoning and tool calling with a 204,800 context. Thinking is off by default on M3 (thinking=adaptive); always-on on M2.7. Latency is not publicly documented on this page.

For self-hosting. Both are open weight with conditional licenses. M3’s Community License allows commercial use with attribution below a $20M revenue threshold; M2.7’s non-commercial license requires prior written authorization for any commercial use. Hardware requirements and quantization are Not publicly documented on this page.

For coding agents. MiniMax-M3 lists coding and publishes SWE-bench Pro (59.0) and Terminal-Bench 2.1 (66.0) plus MCP Atlas (74.2) for tool use; M2.7 lists no coding capability. M3 is more relevant for coding and agent workloads.

For enterprise. Region, SLA, data-residency and compliance terms are not publicly documented — confirm with the vendor. Deployment is comparable (both self-hostable), but the license obligations differ materially: M2.7’s non-commercial boundary is stricter than M3’s community license.

What is uncertain

• Neither model publicly discloses a standalone maximum-output ceiling.
• M2.7’s parameter count is not publicly disclosed.
• The MSA speedup claims (9x prefill / 15x decode) for M3 are vendor claims, not independently measured.
• Benchmark sets do not overlap, and M2.7’s rows use different metrics (ELO, custom end-to-end).

Sources

• MiniMax API platform — model overview (CN)
• MiniMax M3 official blog post
• Hugging Face model card — MiniMax-M3
• Hugging Face model card — MiniMax-M2.7

Labels used above: Official fact (prices, context windows, capabilities and license terms from primary provider sources), Vendor-reported claim (benchmark scores and the MSA speedup claims), 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.