MiniMax M3 vs MiniMax M2.7
MiniMax M3 wins on Overall, Coding, Multimodal. Every numeric field is compared below, with a worked monthly-cost example and a pick rule for each use case.
Verdict at a glance
Choose MiniMax M3 if you…
- Strong agent ability
- Open source
- Low price
- Best for: AI agents, Open source, Chinese
Choose MiniMax M2.7 if you…
- Light and fast
- Open source
- Low price
- Best for: Light tasks, Open source, Low cost
Head-to-head aggregate scores
Scores are 0–100, aggregated from public benchmark information and independently weighted across three leaderboards. Rank is out of 22 tracked models.
Aggregated from public sources and independently weighted; methodology on the Terms page. Scores within 3 points are treated as statistically tied.
Specs & pricing — every field side by side
List API prices in USD per 1M tokens. The highlighted cell is the stronger value/capability on that row.
| Dimension | MiniMax M3 | MiniMax M2.7 | Verdict |
|---|---|---|---|
| Vendor | MiniMax (CN) | MiniMax (CN) | Same vendor |
| Released | 2026.06 | 2026.05 | MiniMax M3 is newer |
| Overall (rank) | 70 · #13 | 64 · #19 | MiniMax M3 +6 |
| Coding | 68 · #13 | 59 · #21 | MiniMax M3 +9 |
| Multimodal | 70 · #18 | 62 · #22 | MiniMax M3 +8 |
| Context window | 1M | 205K | MiniMax M3 larger |
| Max output | 64K | 32K | MiniMax M3 longer |
| Effective-context | 90 | 85 | MiniMax M3 more reliable |
| Input $/1M | $0.6 | $0.27 | MiniMax M2.7 cheaper |
| Output $/1M | $2.4 | $1.08 | MiniMax M2.7 cheaper |
| Cache discount | none | none | Tie |
| Speed | ~60 tok/s | ~70 tok/s | MiniMax M2.7 faster |
| TTFT | 0.5s | 0.4s | MiniMax M2.7 snappier |
| Function calling | 84 | 72 | MiniMax M3 ahead |
| Refusal rate | ~7% | ~6% | MiniMax M2.7 less restrictive |
| English | 70 | 65 | MiniMax M3 |
| Chinese | 88 | 82 | MiniMax M3 |
| Modalities | text | text | Same |
| Open weights | Yes | Yes | Both open |
| Fine-tuning | Yes | Yes | |
| Free tier | Hailuo AI free; open-platform quota | Hailuo AI free | |
| SOC2 / no-train | no / yes | no / yes | |
| Private deployment | Yes | Yes | Both support it |
Fields drawn from vendor public documentation and the Modelspectra 22-model dataset; speed varies with network, concurrency and prompt length. Verify current pricing before purchase.
Dimension-by-dimension analysis
Reasoning & overall intelligence
MiniMax M3 leads the overall aggregate by 6 points (70 vs 64). That is a meaningful, not marginal, edge on hard, multi-step reasoning, while MiniMax M2.7 remains a strong generalist that is not out of its depth on routine work.
Agentic coding
This is a modest gap: MiniMax M3 scores 68 against 59. On multi-file edits, SWE-style tickets and long-horizon agent loops MiniMax M3 needs fewer correction turns; MiniMax M2.7 is still competent for scripts and assisted completion but trails as task complexity rises.
Multimodal
MiniMax M3 leads multimodal 70 vs 62. Neither emits native video, so the comparison is about parsing images and documents, not generation.
Speed & latency
MiniMax M2.7 is faster in interactive use: ~70 tok/s with 0.4s TTFT versus ~60 tok/s with 0.5s TTFT (about 1.2× the throughput). For conversational UIs where perceived responsiveness drives retention, that edge is a real product factor even when raw reasoning is lower.
Context: window vs usable recall
Nominal windows are 1M for MiniMax M3 and 205K for MiniMax M2.7. Effective-context scores point the same way as window size — MiniMax M3 is ahead on usable recall (90 vs 85), so prefer it for long-document work where details cannot be missed.
Price & total cost
MiniMax M2.7 is the cheaper API at $0.27/$1.08 versus MiniMax M3 at $0.6/$2.4 per 1M input/output tokens — list input is about 2.2× lower.
Chinese vs English
English: MiniMax M3 70 vs MiniMax M2.7 65. Chinese: 88 vs 82. For Chinese-language production, MiniMax M3 is the stronger pick. Note that non-Chinese models generally require overseas network access for their APIs.
Tool use & ecosystem
Function-calling score: MiniMax M3 84 vs MiniMax M2.7 72, so MiniMax M3 has the edge on structured tool use. Fine-tuning is available from MiniMax M3 and MiniMax M2.7. Factor in existing SDK/plugin familiarity — switching cost often outweighs a few-point tool-use gap.
Cost worked example — same workload, real token math
Assume a production workload of 100M input + 30M output tokens per month, with 90% of input tokens served from cache. Figures use public list prices.
| Scenario · per month | MiniMax M3 | MiniMax M2.7 | Gap |
|---|---|---|---|
| List priceno cache applied | $132100M in × $0.6 + 30M out × $2.4 | $59100M in × $0.27 + 30M out × $1.08 | 2.24×gap |
| With caching90% of inputs cache-hit | $132no published cache discount | $59no published cache discount | 2.24×gap |
Illustrative model; your input/output mix and cache-hit ratio change the result. Prices are list rates before any enterprise agreement.