As an AI consultancy from Rosenheim, we support companies across the DACH region (Germany, Austria, Switzerland) with GDPR-compliant integration of open-weight models such as MiniMax-M3.1-Flash and MiniMax-M3. Through self-hosting in EU data centres, the models can be deployed for agentic workflows and long-context tasks in compliance with data protection regulations. This overview reflects the state as of October 2026.
Current models at a glance
| Model | Released | Context | Parameters (total / active) | Input | Licence |
|---|---|---|---|---|---|
| MiniMax-M3.1-Flash | 19 October 2026 | 1M tokens | ~428B / 23B | Text, image, video | MiniMax Community License |
| MiniMax-M3 | 1 June 2026 | 1M tokens | ~428B / ~23B | Text, image, video | MiniMax Community License |
| MiniMax-H3 (video) | 31 July 2026 | – | 33B (H3-Base, dense) | Text, image, video, audio | MiniMax H3 Community License (EU excluded) |
All benchmark figures on this page are vendor-reported by MiniMax unless stated otherwise.
MiniMax-M3.1-Flash
On 19 October 2026, MiniMax released M3.1 Flash – initially via the MiniMax API, the M Plan and MiniMax Code. The weights are now available on Hugging Face (MiniMaxAI/MiniMax-M3.1-Flash). The model builds on M3 and keeps its mixture-of-experts architecture with around 428B parameters, 23B of which are active per token. Per the model card it differs from M3 in three ways:
- Sparse attention in every layer: the first three full-attention layers of M3 have been replaced with MSA.
- Quantization-aware training: the routed experts and the attention paths are quantised during training, reducing memory footprint and serving cost.
- DSpark speculative decoding: the draft head ships with the release and replaces the earlier MTP approach.
MiniMax reports, among others, DeepSWE v1.1 73.7%, BankerToolBench 76.2, OSWorld 2.0 61.7, Agents’ Last Exam 29.6 and CyberGym 82.2. Per MiniMax, output speed exceeds 200 tokens/s in everyday use, with a median above 170 tokens/s. Thinking depth is controlled via effort or reasoning_effort in five levels (low, medium, high, xhigh, max; default max). MiniMax has announced a Pro variant of the M3.1 series but has not released it yet.
MiniMax-M3
MiniMax unveiled MiniMax-M3 on 1 June 2026. The model combines coding and agentic capabilities, a 1-million-token context window and native multimodality (image and video understanding); MiniMax guarantees at least 512k tokens of context in the API.
- MSA architecture (MiniMax Sparse Attention): per MiniMax, per-token compute at 1M context drops to 1/20 of the predecessor; prefill is more than 9x and decoding more than 15x faster.
- Native multimodality: per MiniMax, M3 was trained on mixed modalities from the very first step and understands text, images and video in one model – useful for coding with UI screenshots and for document processing.
- Coding benchmarks (per MiniMax): SWE-Bench Pro 59.0%, Terminal-Bench 2.1 66.0%, MCP Atlas 74.2%.
- Reasoning: selectable via the
thinkingparameter (enabled,adaptive,disabled).
The weights are on Hugging Face as MiniMaxAI/MiniMax-M3 (BF16) and MiniMaxAI/MiniMax-M3-MXFP8. Since M3.1 Flash became available, M3 is no longer the first choice on price and efficiency, but it remains available in the API.
MiniMax-H3 (video)
MiniMax-H3 (31 July 2026) is an omni-modal video model that processes text, image, video and audio as a shared context and generates video with native stereo sound. The open H3-Base weights deliver 768p; full 2K quality remains tied to the hosted modules. The MiniMax H3 Community License excludes the EU, the UK, South Korea and the US from the licensed territory, so local operation in the EU is only possible under a separate licence from MiniMax.
Pricing
List prices of the international MiniMax API per 1M tokens (USD):
| Model | Input | Output | Cache read |
|---|---|---|---|
| MiniMax-M3.1-Flash | 0.15 | 0.60 | 0.015 |
| MiniMax-M3 (up to 512k input tokens) | 0.30 | 1.20 | 0.06 |
| MiniMax-M3 (above 512k input tokens) | 0.60 | 2.40 | 0.12 |
For M3, MiniMax shows the prices as a permanent 50% discount on the original list prices (USD 0.60 / 2.40 up to 512k and USD 1.20 / 4.80 above). The priority tier (service_tier: priority) costs 1.5x the standard price. For M3.1 Flash, MiniMax states a single price with no tiering. Third-party providers and cloud platforms set their own prices.
Prices per provider, as of October 2026.
Licence
M3.1 Flash and M3 are not released under MIT, but under the bespoke MiniMax Community License. Per the LICENSE file in each Hugging Face repository, non-commercial use is free. Any commercial use requires visible attribution (‘Built with MiniMax M3.1-Flash’ or ‘Built with MiniMax M3’). If the resulting products or services exceed US$20 million in annual revenue, prior written authorisation from MiniMax is required; below that threshold a one-time email notice is sufficient. For typical use by DACH mid-market companies the revenue threshold usually does not apply, but the attribution and notice obligations apply regardless of revenue and should be reviewed before rollout.
M2.7 is under a Non-Commercial License modelled on MIT (LICENSE file on Hugging Face, MiniMaxAI/MiniMax-M2.7): private use, research and non-commercial purposes are free, but any commercial use – with no revenue threshold – requires prior written authorisation from MiniMax plus the attribution ‘Built with MiniMax M2.7’.
M2, M2.1 and M2.5 are under a Modified MIT licence per their Hugging Face model cards. Anyone wanting to operate commercially without a separate authorisation from MiniMax should have the relevant licence text reviewed beforehand.
Availability in the EU and GDPR
| Model | MiniMax API (storage in the US) | AWS Bedrock | Microsoft Foundry | Google Agent Platform | Self-hosting |
|---|---|---|---|---|---|
| MiniMax-M3.1-Flash | yes | not listed | not listed | not listed | yes |
| MiniMax-M3 | yes | not listed | via Fireworks AI (Global Standard) | not listed | yes (experimental per MiniMax) |
| MiniMax-M2.7 | until 18 Nov 2026, no new users | not listed | not listed | not listed | yes (Non-Commercial License) |
| MiniMax-M2.5 | until 18 Nov 2026, no new users | yes, In-Region incl. Frankfurt, Ireland, Stockholm | not listed | not listed | yes |
| MiniMax-M2 | shut down | legacy until 30 Mar 2027, In-Region | not listed | retired on 21 Oct 2026 | yes |
Who processes the data?
Per its privacy policy (updated 30 March 2026), the international platform (platform.minimax.io, api.minimax.io) is operated by Nanonoble Pte. Ltd., registered in Singapore, as the controller. Per the policy, personal data is stored in a data centre in the United States. MiniMax refers to a cloud provider certified under the EU-US Data Privacy Framework and to standard contractual clauses for transfers to group companies and third parties. The Chinese platform (platform.minimaxi.com) has its own terms.
There is no European Commission adequacy decision for Singapore. Anyone using the API with personal data therefore needs an assessment under Art. 44 et seq. GDPR: a contracting party in Singapore, storage in the US, standard contractual clauses, a transfer impact assessment and a data processing agreement under Art. 28 GDPR. How the cloud provider’s DPF certification affects this transfer chain has to be assessed case by case.
Note: This is a factual summary and does not constitute legal advice. Whether a specific deployment is permissible depends on your role, the categories of data, contractual arrangements and safeguards in the individual case.
Routes for EU deployment
- Self-hosting: Full data control on your own hardware or on GPU instances in hyperscaler EU regions (e.g. AWS Frankfurt and Ireland, Azure West Europe and Germany West Central, Google Cloud Frankfurt and Belgium) as well as in your own data centre.
- AWS Bedrock carries MiniMax M2, M2.1 and M2.5 as native models. Per the AWS model cards, only In-Region inference is available (no Geo or Global cross-region routing) – including eu-central-1 (Frankfurt), eu-west-1 (Ireland), eu-south-1 (Milan) and eu-north-1 (Stockholm); eu-west-2 (London) is in the United Kingdom. M2.5 is active, while M2 and M2.1 are in legacy status with end of life on 30 March 2027. M2.7, M3 and M3.1 Flash are not listed on Bedrock.
- Microsoft Foundry carries MiniMax M3 as a Fireworks on Foundry offering (pay-per-token as Global Standard). Per Microsoft, data is shared with Fireworks AI, data zones are available in the US only, and Fireworks on Foundry is currently excluded from EU Data Boundary commitments.
- Google Agent Platform: Google retired the MaaS endpoint for MiniMax M2 on 21 October 2026; per Google, all MiniMax MaaS endpoints are deprecated or retired. Self-deployment from Model Garden remains possible.
- Other inference providers: Together AI, Fireworks AI, OpenRouter. Check regions, retention and terms directly with the provider.
Hardware for self-hosting
With MoE models, all expert weights have to be loaded – 23B active parameters do not mean that memory is only needed for those.
- MiniMax-M3: per MiniMax’s documentation, self-hosting is still experimental (SGLang development image). The reference is a node with 8x NVIDIA B200 and the MXFP8 weights (around 444 GB); validated configurations also include 4x B300 with MXFP8 and 8x H200 with BF16 weights (around 854 GB).
- MiniMax-M3.1-Flash: the quantisation-aware trained weights are considerably smaller than M3’s. MiniMax provides recipes for SGLang, vLLM and TokenSpeed but does not name reference hardware in the model card.
- MiniMax-M2.7: per MiniMax four high-memory GPUs (tensor parallelism 4), with no official minimum.
Migration and breaking changes
- M2.5 and M2.7 to M3.1 Flash: since 19 October 2026 MiniMax no longer accepts new users for either model family, and both will be shut down on 18 November 2026, after which API calls fail. In the M Plan and the Token Plan, MiniMax has automatically routed requests for M2.5 and M2.7 to M3.1 Flash since 19 October 2026; the existing model names keep working there.
- Reasoning: M3.1 Flash thinks by default at level
max; depth is controlled viaeffortorreasoning_effort. On the OpenAI-compatible protocol, thinking content is returned separately inreasoning_content. M3 uses thethinkingparameter instead (enabled,adaptive,disabled). - Cloud platforms: on AWS Bedrock, M2 and M2.1 end on 30 March 2027; the Google MaaS endpoint for M2 has been retired since 21 October 2026.
Model lifecycle
| Model | Released | Status in the MiniMax API | Other platforms |
|---|---|---|---|
| MiniMax-M3.1-Flash | 19 October 2026 | current | – |
| MiniMax-M3 | 1 June 2026 | available | Microsoft Foundry (Fireworks) |
| MiniMax-M2.7, M2.7-highspeed | 18 March 2026 | no new users since 19 Oct 2026, shutdown on 18 Nov 2026 | – |
| MiniMax-M2.5, M2.5-highspeed | February 2026 | no new users since 19 Oct 2026, shutdown on 18 Nov 2026 | AWS Bedrock active |
| MiniMax-M2.1 | 22 December 2025 | shut down | AWS Bedrock legacy until 30 Mar 2027 |
| MiniMax-M2 | 27 October 2025 | shut down | AWS Bedrock legacy until 30 Mar 2027; Google MaaS retired on 21 Oct 2026 |
The weights of M2, M2.1, M2.5 and M2.7 remain available on Hugging Face (MiniMaxAI) for self-hosting.
Local or Cloud? Calculate the Cost
For agentic coding workloads, a precise cost calculation between API usage and self-hosting often pays off. The local-vs-cloud AI inference calculator from ai-prices.eu lets you compare hardware, electricity and operating costs against cloud API pricing and determine the break-even point for your workload.
Our Recommendation
For new projects: MiniMax-M3.1-Flash, self-hosted in the EU. It is newer and cheaper than M3, comes with smaller weight files and keeps the 1M-token context as well as image and video input. Review the attribution and notice obligations of the MiniMax Community License before rollout.
For existing M3 setups: M3 remains available in the API. Switching to M3.1 Flash pays off mainly for price and hardware requirements; MiniMax itself still classes M3 self-hosting as experimental.
On AWS with a fixed EU region: MiniMax-M2.5 via Bedrock In-Region (e.g. Frankfurt, Ireland, Stockholm) under a Modified MIT licence – an older, text-only generation. M2 and M2.1 only run there until 30 March 2027.
Existing M2.5 and M2.7 users of the MiniMax API should complete their migration before 18 November 2026.
Not recommended for personal or confidential data: the direct MiniMax API without an assessment under Art. 44 et seq. GDPR (contracting party in Singapore, storage in the US) and Fireworks on Foundry while it is excluded from the EU Data Boundary.
We support you with:
- Infrastructure planning and hardware sizing (especially for 1M-context workloads)
- Deployment, quantisation and optimisation
- Integration into existing agent frameworks
- Compliant usage within CompanyGPT
- Fine-tuning for specific use cases
MiniMax-M3.1-Flash and M3 combine a 1M-token context, open weights and native multimodality – provided the infrastructure and licence questions are settled.
