MiniMax's current open-weight flagship is MiniMax M3, released on June 1, 2026. Its most recent release is M3.1-Flash-Preview, which went live on September 27, 2026 inside the MiniMax Code tool and the company's Token Plan subscriptions. A full M3.1 with open weights and a per-token API price has not been released, and MiniMax has not given a date for one. This page lists every MiniMax text model from the abab series to the M series, with dates, context windows, and whether the weights are open.
MiniMax Model Timeline
| Model | Released | Context window | Notable change | Status |
|---|---|---|---|---|
| abab1 | April 2022 | Not disclosed | First MiniMax text model | Superseded |
| abab2, abab3 | June and October 2022 | Not disclosed | abab3 raised the parameter count | Superseded |
| abab6 | January 2024 | Not disclosed | First mixture-of-experts model in the series, per secondary sources | Superseded |
| abab6.5, abab6.5s | April 17, 2024 | 200K | MiniMax described abab6.5 as a trillion-parameter model. abab6.5s is the faster variant | Superseded |
| MiniMax-Text-01, MiniMax-VL-01 | January 2025 | 1M in training, up to 4M at inference | Open weights. 456B parameters, 45.9B active per token | Superseded |
| MiniMax-M1 | June 2025 | 1M | Open-weight reasoning model under Apache 2.0, with 40K and 80K thinking budgets | Superseded |
| MiniMax-M2 | October 2025 | 204,800 | 230B parameters, 10B active. Built for coding and agents at $0.30 input and $1.20 output | Legacy, still on the API |
| MiniMax-M2.1 | About December 2025 | 204,800 | Coding update to M2 | Legacy, still on the API |
| MiniMax-M2.5, M2.5-Lightning | February 12, 2026 | 204,800 | MiniMax reported 80.2 percent on SWE-Bench Verified | Legacy, still on the API |
| MiniMax-M2.7 | March 18, 2026 | 204,800 | Agent workflows and office tasks. Weights reached Hugging Face in April | Available |
| MiniMax-M3 | June 1, 2026 | 1M | About 428B parameters, 23B active. Native image and video input, computer use | Current open-weight flagship |
| M3.1-Flash-Preview | September 27, 2026 | 1M | Five effort levels from low to max. Thinking cannot be switched off | Preview, closed |
Two things stand out. MiniMax reached a 1M-token window with the 01 series and M1, dropped to 204,800 tokens for the whole M2 line, and returned to 1M with M3. And the naming changed twice: abab until 2024, the 01 series for one release, then M1 onward. MiniMax listed on the Hong Kong Stock Exchange on January 9, 2026, between M2.1 and M2.5.
Open or Closed, and API Price
| Model | Weights | API price per million tokens (input / output) |
|---|---|---|
| abab series | Closed, API only | Not on the current price list |
| MiniMax-Text-01 | Open, under MiniMax's own model agreement | Not on the current price list |
| M1 | Open, Apache 2.0 | Not on the current price list |
| M2 | Open, modified MIT | $0.30 / $1.20 |
| M2.1, M2.5, M2.7 | Open on Hugging Face | $0.30 / $1.20. High-speed variants cost $0.60 / $2.40 |
| M3 | Open, minimax-community licence | $0.30 / $1.20 for prompts up to 512K tokens, $0.60 / $2.40 above |
| M3.1-Flash-Preview | Closed. No model card | No per-token price. Token Plan subscriptions at $22, $55, and $132 per month, as reported by DataNorth |
The base price has not moved since M2. Every M-series model on MiniMax's price list costs $0.30 input and $1.20 output per million tokens, so an older model is not cheaper to keep running, and M3 adds a surcharge only for very long prompts.
What Is Next
M3.1. MiniMax has not announced a date for a full M3.1 release, an API model ID, or open weights. DataNorth, which covered the preview on September 28, notes that MiniMax has a record of shipping the full model a few weeks after a preview. That is the outlet's observation, not a MiniMax commitment. Codersera's tracker recorded no M3.1 and no M3 Pro as of August 31, 2026, so the preview is the first M3.1-branded model anyone can use.
Open weights for M3.1. Unknown. Every M-series model from M1 to M3 was released with weights. The preview is the first M-series release without them, and MiniMax has not said whether that will change.
We found no MiniMax statement about an M4. Any date you see for one is a guess.
Brand Visibility Implications
- Open weights run where nobody can watch. M2 through M3 can be self-hosted, so they answer questions inside companies and developer tools with no retrieval layer and no public log. What the model learned about a brand at training time is what those users get.
- MiniMax does not publish a knowledge cutoff on the M3 model card. You cannot look up what the model should know. The only way to find out what it recalls about a brand is to ask it.
- Old versions stay in service. Four M2-series models remain on the API at the same price as M3. A product built on M2.5 in February has no cost reason to upgrade, so its picture of your brand stays where it was.
- Each release is a new baseline. M2.7 to M3 changed the attention architecture and raised the context window from 204,800 tokens to 1M. Compare answers within a version, and re-run them when the version changes. See how training cutoffs change brand recall.
For how MiniMax sits among the other Chinese open-weight labs, see the Chinese open-source LLM comparison and the DeepSeek model lineage.
Methodology
This page is built from vendor documentation and dated reporting, not from Presenc AI measurements. Primary sources are used wherever one exists: OpenAI's model pages and deprecations page, Google's Gemini API changelog and deprecations page, xAI's release notes and model list, DeepSeek's API changelog and pricing page, and MiniMax's pricing page and Hugging Face model cards. Older release dates that vendors no longer document come from the relevant Wikipedia articles. Where a figure comes only from a secondary tracker or a single outlet, the page names it. "Not disclosed" means the vendor material we read does not state the figure. Status as of October 1, 2026.
How Presenc AI Helps
Presenc AI tracks how AI models describe and recommend your brand, version by version, so you can see what changed when a vendor ships a new model. That includes open-weight models that are deployed far beyond the vendor's own app.