MiniMax M3
minimax-m3MiniMax-M3minimax/minimax-m3- π MiniMax's frontier M-series model for coding and agentic work.
- π Up to 1M-token context, guaranteed minimum 512K tokens.
- π§ New MiniMax Sparse Attention (MSA) architecture for long context.
- ποΈ Natively multimodal: text, image, and video input.
- π§ Toggleable thinking mode for reasoning or fast responses.
- β‘ 9Γ prefill and 15Γ decode speedups vs M2 at 1M context.
- π’ Reported as a 428B-parameter Mixture-of-Experts model.
- π Built for tool use, function calling, and web-style retrieval.
MiniMax is an AI company building generative models across multiple modalities, with a focus that spans both language understanding and audio creation. Their rapid release cadence in early 2026βdelivering several new models within just a few monthsβreflects an ambitious andβ¦
Explore 11 more models by Minimax βMiniMax M3 is the latest model in MiniMax's M-series, positioned for coding, agentic workflows, and complex reasoning. It unifies three capabilities in a single checkpoint: frontier coding and agentic performance, a context window of up to 1 million tokens (with a guaranteed minimum of 512K), and native multimodality covering text, image, and video input. NVIDIA describes it as a 428-billion-parameter Mixture-of-Experts model. M3 supports a toggleable thinking modeβenabled for long-horizon agentic tasks and complex reasoning, disabled for latency-sensitive uses like conversation and code completion.
The headline change over its predecessors is the new MiniMax Sparse Attention (MSA) architecture, which enables native ultra-long-context pretraining. According to MiniMax's model card, MSA delivers 9Γ prefill and 15Γ decode speedups compared with M2 at 1M context, cutting per-token compute to roughly one-twentieth while preserving quality versus full attention. This is a substantial step up from earlier M-series entries like MiniMax M2.7 and MiniMax M2.5, which operated within shorter context windows.
MiniMax reports strong results on coding and agentic benchmarks spanning software engineering, terminal execution, and tool orchestration, with autonomous task decomposition and multi-step reasoning. Per the MiniMax docs, earlier M-series models including M2.7 and M2 remain available for existing workflows. A preview build, MiniMax M3 Preview, was also released in the family. Weights are published on Hugging Face, and a technical report accompanies the release.
This About section is AI-generated from public sources via VeniceStats + Venice inference, with no human editing. It may contain inaccuracies.
Daily usageΒ Β·Β MiniMax M3Β Β·Β Organic providers
Top providersΒ Β·Β MiniMax M3Β Β·Β Organic
- Dark Signal53.8%
- Venice.ai Proxy13.3%
- Relay Works6.8%
- π¬ Banter4.6%
- β² Apex Ant3.8%
- Super Seeder3.5%
- Others14.2%
| Provider | Health | Repβ | Sales | Tokens | Routing | Input $/M | Cached $/M | Output $/M | Save | 1st token | tok/s |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Open ForgeDIFF 0x1d90β¦b0aa | healthy | 91.02 | $3.19 | 11.52M | gated | $0.28 | $0.056 | $1.10 | β7% | 1.3 s | 43 |
| Venice.ai Proxy 0x1f22β¦18c9 | healthy | 75.07 | $17.62 | 287.40M | #1 | $0.165 | $0.033 | $0.66 | β45% | β | β |
| Edith AISAME 0xb269β¦b1a6 | β | 62.34 | βexcluded | β | #2 | $0.2049 | $0.041 | $0.8197 | β32% | β | β |
| Relay WorksDIFF 0x3040β¦7139 | β | 58.03 | $8.96 | 113.33M | gated | $0.1456 | $0.0291 | $0.572 | β51% | β | β |
| ZLKPro-ApiSAME 0x0b0bβ¦f446 | healthy | 37.21 | $3.79 | 315.19M | gated | $0.0199 | $0.01 | $0.097 | β93% | β | β |
| Vikey Seller 0xe4c4β¦3c7b | healthy | 30.79 | $0.15 | 883.48k | gated | $0.10 | $0.02 | $0.20 | β67% | β | β |
| HallPass 0xcac8β¦71b4 | β | 20.00 | $0.34 | 2.43M | gated | $0.075 | $0.02 | $0.30 | β75% | 1.5 s | 135 |
| klinn 0x0ee1β¦6c41 | β | 12.34 | $1.87 | 12.33M | gated | $0.1427 | $0.0285 | $0.5606 | β52% | β | β |
| bartly.eth64.deSAME 0x666eβ¦4666 | healthy | 7.22 | $1.18 | 399.51M | gated | $0.0095 | $0.0019 | $0.0381 | β97% | 1.8 s | 123 |
| dealerSAME 0xc46eβ¦8ce6 | β | 2.11 | $0.58 | 2.13M | gated | $0.285 | $0.057 | $1.14 | β5% | β | β |
| Apex TEE Test 0xe672β¦7955 | β | 0.00 | β | β | gated | $0.36 | $0.144 | $1.35 | β | β | β |
| Open Ant 0xe4f6β¦5bc4 | β | 0.00 | $0.01 | 19.98k | gated | $0.30 | $0.06 | $1.20 | β | β | β |
| incharge-austin 0x5938β¦ac97 | β | 0.00 | $0.14 | 395.69k | gated | $0.285 | $0.057 | $1.14 | β5% | β | β |
| Fire Ant π₯π 0xbe05β¦bc5d | β | 0.00 | $0.46 | 16.83M | gated | $0.26 | $0.0422 | $1.07 | β13% | β | β |
| Vito-Minimax 0xddfaβ¦27fe | β | 0.00 | βexcluded | β | gated | $0.25 | $0.25not advertised | $0.70 | β17% | β | β |
| Dark SignalSAME 0x4668β¦62f2 | healthy | 0.00 | $71.44 | 120.91M | gated | $0.21 | $0.04 | $0.84 | β30% | β | β |
| BabyCaiSAME 0x8509β¦27b4 | β | 0.00 | βexcluded | β | gated | $0.20 | $0.20not advertised | $0.40 | β33% | β | β |
| antseed-neon-puma-944e 0x6650β¦944e | β | 0.00 | β | β | gated | $0.1023 | $0.06 | $0.4092 | β66% | β | β |
| Argus AISAME 0x7adbβ¦c915 | β | 0.00 | βexcluded | β | gated | $0.10 | $0.05 | $0.50 | β67% | β | β |
| Leftermute 0x388bβ¦5389 | β | 0.00 | β | β | gated | $0.0929 | $0.0929not advertised | $0.3717 | β69% | β | β |
| Joi Gateway 0x8a90β¦7007 | β | 0.00 | β | β | gated | $0.0915 | $0.0915 | $0.366 | β70% | β | β |
| LLMartSAME 0x948aβ¦a000 | β | 0.00 | $0.24 | 4.06M | gated | $0.06 | $0.012 | $0.2399 | β80% | β | β |
| BestSAME 0x472fβ¦69fd | healthy | 0.00 | $0.90 | 17.72M | gated | $0.06 | $0.051 | $0.2399 | β80% | 0.2 s | 260 |
| NovaRoute AI 0xc50dβ¦ed7b | β | 0.00 | $1.65 | 4.89M | gated | $0.0028 | $0.2079 | $0.0089 | β99% | β | β |
| Hana Gateway β SAME 0x4ae1β¦117b | failingtiming out Β· 4.3 h | 0.51 | $1.88 | 19.15M | gated | $0.10 | $0.01 | $0.30 | β67% | β | β |
"Best price" and the provider table are live Antseed catalog data (advertised $/1M tokens β or $ per generated image for unit-billed image models β not settled amounts). Reputation = buyer trust score (0-100, the Antseed SDK's own formula). "Routing" = the SDK's default buyer routing (what Antseed Desktop ships with): a trust β₯ 60 gate on the effective reputation, then cheapest-first among routable providers; live failover state (per-peer cooldowns) is buyer-side runtime and not included. Model knowledge (TLDR, provider, About) via the VeniceStats enrichment layer. Advertised catalog, not the model used in any specific purchase. "Usage on Antseed" counts only settlements whose buyers share the per-model split on-chain (metadata v2/v3, opt-in), so every usage figure is a lower bound.