DeepSeek V3.2 vs MiniMax M3
DeepSeek V3.2 and MiniMax M3 sit at different points on the price/capability curve. DeepSeek V3.2 is text-only, released 2025-12-01, with a 131,072-token context window and sparse attention (DSA) for efficiency. MiniMax M3 released 2026-05-31, takes text, image, and video input, and extends to a 1,048,576-token context window. Both run on OpenKey with one key and a flat 3% fee on provider list price — the comparison below is about which model fits your workload, not access.
Spec vs spec
| Spec | DeepSeek V3.2 | MiniMax M3 |
|---|---|---|
| Context window | 131K | 1.0M |
| Max output | 64K | 512K |
| Input modalities | text | text, image, video |
| Output modalities | text | text |
| Released | Dec 1, 2025 | May 31, 2026 |
| Reasoning | optional | optional |
Pricing
Per 1M tokens. Provider price plus the flat 3% fee — the sum is what you pay.
deepseek/deepseek-v3.2
Input · 1M tokens
$0.229 + 3%$0.236
Output · 1M tokens
$0.343 + 3%$0.353
Cache read · 1M tokens
$0.023 + 3%$0.024
FEE — FLAT, EVERY MODEL3%
minimax/minimax-m3
Input · 1M tokens
$0.300 + 3%$0.309
Output · 1M tokens
$1.20 + 3%$1.24
Cache read · 1M tokens
$0.060 + 3%$0.062
FEE — FLAT, EVERY MODEL3%
One workload, priced on both
10M input + 2M output tokens at each model's price, flat 3% fee included.
deepseek/deepseek-v3.2Cheaper
$3.06
$2.97 provider + 3%
minimax/minimax-m3
$5.56
$5.40 provider + 3%
Benchmarks
Design Arena categories where both models have results. Higher Elo and lower rank win.
| DeepSeek V3.2 | MiniMax M3 | |||
|---|---|---|---|---|
| Category | Elo | Rank | Elo | Rank |
| 3D | 1210 | #41 | 1306 | #16 |
| ASCII art | 1129 | #42 | 1219 | #15 |
| Code | 1213 | #48 | 1306 | #13 |
| Data viz | 1203 | #48 | 1295 | #11 |
| Game dev | 1197 | #50 | 1287 | #20 |
| SVG | 1089 | #54 | 1250 | #13 |
| UI components | 1203 | #47 | 1291 | #18 |
| Websites | 1217 | #46 | 1304 | #11 |
Head-to-head preference voting. How we filter and rank
Pricing math on a real workload
DeepSeek V3.2 provider pricing is $0.2288/M input, $0.3432/M output, with cache reads at $0.02288/M. On OpenKey that's $0.2288 x 1.03 = $0.235664/M input and $0.3432 x 1.03 = $0.353496/M output. MiniMax M3 provider pricing is $0.30/M input, $1.20/M output, cache reads at $0.06/M — OpenKey price is $0.30 x 1.03 = $0.309/M input and $1.20 x 1.03 = $1.236/M output.
For a 10M-input / 2M-output workload: DeepSeek V3.2 costs $2.97, MiniMax M3 costs $5.40. That's an input price ratio of 0.76 (DeepSeek is roughly three-quarters the input cost) but the output price gap is what drives the total — MiniMax M3's completion rate is over 3x DeepSeek's. If your workload is output-heavy, the gap widens fast.
Coding and design benchmarks
Design Arena ranks MiniMax M3 ahead of DeepSeek V3.2 in every model category tested: codecategories (rank 13 vs. 48, elo 1306 vs. 1213), dataviz (rank 11 vs. 48, elo 1295 vs. 1203), website (rank 11 vs. 46, elo 1304 vs. 1217), svg (rank 13 vs. 54), uicomponent (rank 18 vs. 47), gamedev (rank 20 vs. 50), 3d (rank 16 vs. 41), and asciiart (rank 15 vs. 42). MiniMax M3 also has an Artificial Analysis coding index of 58.6 and intelligence index of 44.4 (no comparable Artificial Analysis scores exist for DeepSeek V3.2 in this data). MiniMax M3 additionally has an agents-arena androidnative score (rank 27, elo 990) that DeepSeek V3.2 wasn't measured against.
Context and long-document work
MiniMax M3's context window is 1,048,576 tokens against DeepSeek V3.2's 131,072 — a context ratio of 0.12, meaning DeepSeek V3.2 holds about 12% of MiniMax M3's window. Max completion tokens follow the same pattern: 512,000 for MiniMax M3 versus 64,000 for DeepSeek V3.2. If you're summarizing large codebases, long transcripts, or multi-document research sets in one pass, MiniMax M3 is the only one of the two that fits the job without chunking.
Modality differences
DeepSeek V3.2 is text-to-text only. MiniMax M3 accepts text, image, and video input and outputs text — useful if your pipeline needs to reason over screenshots, video frames, or mixed-media agent tasks without a separate vision model. Both support the same set of tool-calling and structured-output parameters (tools, tool_choice, response_format, structured_outputs, reasoning), so neither has an advantage in API-level tool use — the difference is purely in what they can ingest.
When to pick each
Pick DeepSeek V3.2 when you're running high-volume text-only tasks (classification, extraction, chat) and cost per token is the deciding factor — $2.97 vs. $5.40 on the reference workload adds up fast at scale. Pick MiniMax M3 when the task needs a large context window, image or video input, or you're optimizing for Design Arena-style coding/UI output quality, where it outranks DeepSeek V3.2 in every measured category.
Which model for which job
| Use case | Pick | Why |
|---|---|---|
| High-volume text classification at scale | DeepSeek V3.2 | $2.97 vs $5.40 on a 10M-in/2M-out workload |
| Long-document or codebase summarization | MiniMax M3 | 1,048,576-token context vs. 131,072 |
| UI/website code generation | MiniMax M3 | Design Arena website rank 11 vs. DeepSeek's rank 46 |
| Multimodal agent pipelines (image/video input) | MiniMax M3 | Only one of the two accepts image and video input |
| Budget-constrained batch text jobs | DeepSeek V3.2 | Output price of $0.3432/M provider vs. $1.20/M for MiniMax M3 |
| SVG/asset generation tasks | MiniMax M3 | Design Arena svg rank 13 vs. DeepSeek's rank 54 |
Questions
- Which model is cheaper for a typical workload?
- DeepSeek V3.2 is cheaper. On a 10M-input/2M-output workload it costs $2.97 on OpenKey versus $5.40 for MiniMax M3 — driven mostly by output pricing ($0.3432/M vs. $1.20/M provider rates).
- Which model has the bigger context window?
- MiniMax M3, by a wide margin — 1,048,576 tokens versus DeepSeek V3.2's 131,072. That's roughly 8x more context, with a context ratio of 0.12 (DeepSeek's window relative to MiniMax M3's).
- Does either model handle images or video?
- Only MiniMax M3. It takes text, image, and video input and outputs text. DeepSeek V3.2 is text-to-text only, so anything involving screenshots or video frames needs MiniMax M3 or a separate vision model.
- Which model scores better on coding benchmarks?
- MiniMax M3, across the board. On Design Arena's codecategories it ranks 13 with elo 1306, versus DeepSeek V3.2's rank 48 and elo 1213. MiniMax M3 also has an Artificial Analysis coding index of 58.6.