Last verified: November 9, 2026 (model status, prices, context windows, API parameters, and code examples); benchmark, architecture, and license details July 18, 2026.
MiniMax closed M2.7 to new users on October 19, 2026, and set November 18, 2026 as its retirement date. After that date, API requests to MiniMax-M2.7 and MiniMax-M2.7-highspeed no longer work, so pay-as-you-go integrations need a new model ID. On M Plan and Token Plan, MiniMax has routed requests that name M2.5 or M2.7 to M3.1 Flash since October 19, 2026. The old model names keep working for backward compatibility, and plan usage allowances are unchanged.[17]
That turns the MiniMax M3 vs M2.7 question into a migration choice between the two text models on MiniMax’s Model List, MiniMax-M3.1-Flash and MiniMax-M3. Both have a 1,000,000-token context window and accept text, images, and video. M3.1 Flash costs half of M2.7’s last published standard rates ($0.15 versus $0.30 per million input tokens, $0.60 versus $1.20 output) and tunes its always-on thinking with five effort levels. M3 matches M2.7’s last rates up to 512K input tokens and is the better fit when you need its Priority tier or thinking you can switch off.
How to read the numbers: Benchmark figures below are MiniMax-reported, not tests run by this website. M2.7 prices are the last rates MiniMax published; its current price list no longer includes them. Prices, access rules, and license terms are dated snapshots; verify the official sources before purchasing capacity or deploying commercially.
M2.7 retirement: who is affected and where to move
MiniMax’s “MiniMax M2.5 and M2.7 Model Retirement Notice,” published in its Model List, covers four hosted model IDs: MiniMax-M2.5, MiniMax-M2.5-highspeed, MiniMax-M2.7, and MiniMax-M2.7-highspeed. What the October 19 and November 18, 2026 dates mean depends on how you call the model.[17]
| How you call M2.7 | What changes | What to do |
|---|---|---|
| Pay-as-you-go API with a standard API Key | Closed to new users since October 19, 2026. After November 18, 2026, requests that name MiniMax-M2.7 or MiniMax-M2.7-highspeed fail. | Change the model ID to MiniMax-M3.1-Flash or MiniMax-M3 after a regression run on the new model. |
| M Plan or Token Plan with a Subscription Key | Since October 19, 2026, MiniMax has routed requests that name M2.5 or M2.7 to M3.1 Flash. The old names keep working for backward compatibility, and plan usage allowances are unchanged. | No code change is needed. Re-test the outputs you rely on, and consider naming MiniMax-M3.1-Flash directly so logs and effort settings match the model that answers. |
| Self-hosted open weights | Not covered: the notice concerns MiniMax’s hosted API. The MiniMaxAI/MiniMax-M2.7 repository is on Hugging Face. | MiniMax’s “Deploy MiniMax-M2.7” guide covers existing M2.7 workflows and points new projects that need native multimodal input or a 1M-token context to M3.[19] |
MiniMax’s Model List names only MiniMax-M3.1-Flash and MiniMax-M3 under its available LLMs. Its Retired Models table shows MiniMax-M2.7 and MiniMax-M2.7-highspeed together with the M2.5, M2.1, M2, and M2-her entries, all without dates.[17] Our M2.7 retirement and migration guide lists what to change in pay-as-you-go requests.
Where to move: M3.1 Flash or M3
| Area | MiniMax M3.1 Flash | MiniMax M3 | M2.7, for reference |
|---|---|---|---|
| Model ID | MiniMax-M3.1-Flash | MiniMax-M3 | MiniMax-M2.7 and MiniMax-M2.7-highspeed, both in the Retired Models table (retirement date November 18, 2026) |
| Input / output, per 1M tokens | $0.15 / $0.60 at any prompt length | $0.30 / $1.20 up to 512K input; $0.60 / $2.40 above | $0.30 / $1.20; highspeed $0.60 / $2.40 (last published rates) |
| Cache read, per 1M tokens | $0.015 | $0.06 up to 512K input; $0.12 above | $0.06 (last published rate) |
| Context window | 1,000,000 tokens | 1,000,000 tokens | 204,800 tokens |
| Thinking | Always on; effort low, medium, high, xhigh, or max (default max) | Adaptive or disabled; the default differs by API format; no effort control | Always on |
| Image and video input | Yes | Yes | No |
| Priority tier | Not supported | service_tier: "priority", at 1.5 times the Standard price | Not listed |
For most M2.7 workloads, begin the evaluation with M3.1 Flash. It is the model that answers plan requests naming M2.7, it costs half of M2.7’s last standard input and output rates, and it always thinks, as M2.7 did. Set its effort explicitly, because the default is max. Choose M3 when you need the Priority tier or calls with thinking switched off. If you relied on M2.7-highspeed for speed, measure M3.1 Flash’s latency on your own traffic; its rates are 75% below highspeed’s last published rates, and M3 Priority is MiniMax’s documented option for priority admission.[5][18]
MiniMax M3 vs M2.7: quick comparison
| Area | MiniMax M3 | MiniMax M2.7 |
|---|---|---|
| Release date | June 1, 2026 | March 18, 2026 |
| Hosted API status | Listed under Available Models, with M3.1 Flash | Closed to new users on October 19, 2026; retirement date November 18, 2026 |
| Hosted model ID | MiniMax-M3 | MiniMax-M2.7; faster hosted variant MiniMax-M2.7-highspeed; both in the Retired Models table |
| Combined context window | 1,000,000 tokens | 204,800 tokens |
| Documented output setting | 131,072 recommended; 524,288 hard maximum, subject to the combined context budget | 65,536 recommended; 204,800 hard maximum, subject to the combined context budget |
| Compatible inputs | Text, images, video, tool calls, tool results, and thinking content | Text, tool calls, tool results, and thinking content; no image or video input through the compatible text APIs |
| Output modality | Text | Text |
| Reasoning behavior | Adaptive or disabled; endpoint defaults differ | Thinking cannot be disabled in the documented compatible APIs |
| Architecture | Multimodal MoE with MiniMax Sparse Attention; approximately 428B total and 23B activated parameters | Text-focused, full-attention MoE; Hugging Face reports 229B total parameters |
| Public weights | MiniMaxAI/MiniMax-M3 | MiniMaxAI/MiniMax-M2.7 |
| Standard API input / output price | $0.30 / $1.20 per million tokens at up to 512K input; $0.60 / $2.40 above 512K input | $0.30 / $1.20 per million tokens, the last rates MiniMax published |
| Weight-license commercial path | Conditional commercial use: attribution plus notice below the stated revenue threshold; prior authorization above it | Prior written MiniMax authorization required for any commercial use of the weights |
| Where it fits after the retirement | Long, multimodal, agentic, or repository-scale work, and calls that need the Priority tier or thinking switched off | Self-hosted M2.7 workflows, and a historical baseline for comparisons |
For deep specifications rather than a head-to-head decision, use the separate MiniMax M3.1 Flash guide, MiniMax M3 model guide, and MiniMax M2.7 model guide. This page concentrates on the differences that matter when an M2.7 workload moves to M3 or M3.1 Flash.
On this page
- Release and product positioning
- Coding and agent workflows
- Context and output limits
- Multimodality and reasoning
- Architecture and open weights
- Benchmark claims
- API pricing, M Plan, and Token Plan
- License differences
- Decision matrix
- Migration checklist: M2.7 to M3 or M3.1 Flash
- Frequently asked questions
Release sequence and product positioning
MiniMax released M2.7 on March 18, 2026 and described it as its first model to participate deeply in parts of its own development process. The phrase “self-improvement” refers to an internal development workflow in which the model helped maintain memory, create skills, analyze failed experiments, and revise an agent scaffold. It did not mean that an ordinary M2.7 API request retrained or rewrote the production model.[2]
M3 followed on June 1, 2026. MiniMax positioned it around three combined capabilities: coding and agent work, a one-million-token context window, and native image/video understanding. MiniMax Code also moved to M3, and MiniMax revised its Token Plan tiers around M3 usage.[1]
M3.1 Flash followed on October 19, 2026, the same day MiniMax closed the M2.5 and M2.7 families to new users. MiniMax set November 18, 2026 as their retirement date. On this page, M2.7 is therefore a historical baseline: its release claims, benchmarks, and architecture describe what you are migrating from, not an option for new API work.[17][20]
Coding and agent workflows
All three models are designed for more than single-turn code generation. They can participate in an agent loop that reads repository files, calls tools, runs tests, inspects errors, and revises a patch. In every case, the model does not obtain shell, browser, database, or desktop access by itself. The surrounding agent harness supplies those capabilities and must enforce permissions.
Where M3 and M3.1 Flash change the workflow
- Larger working history: M3 and M3.1 Flash can retain substantially more code, specifications, tool results, and conversation history in one combined context budget.
- Visual evidence: screenshots, diagrams, interface recordings, and video can be supplied alongside text through supported compatible requests.
- Controllable reasoning: on M3, an application can use adaptive thinking for difficult work and disable it for low-latency extraction or completion. M3.1 Flash always thinks, but its effort level sets how deeply.
- Long-horizon positioning: MiniMax’s M3 release emphasizes extended tool sessions, repository-scale work, and desktop-oriented agents.
Old reasons to keep M2.7, and what replaces them
- Prompts, parsers, and tool loops tuned to M2.7: that tuning does not carry over automatically. Rerun your regression suite on M3.1 Flash or M3 before moving production traffic. On M Plan and Token Plan, requests that name M2.7 are answered by M3.1 Flash, so re-test there too.
- Text-only work within 204,800 tokens: it fits inside the 1,000,000-token window of either replacement with room to spare, so context is not a migration blocker.
- The
MiniMax-M2.7-highspeedvariant: MiniMax lists no highspeed ID for M3 or M3.1 Flash. Measure M3.1 Flash’s latency on your own traffic before paying for M3’s Priority tier. - Exact reproduction of an M2.7 result: after November 18, 2026, the hosted API no longer serves M2.7. The closest rerun is on the self-hosted open weights, and hosted and self-hosted results can differ.
For the official coding product rather than a custom harness, see the MiniMax Code setup guide. For a route chooser covering Code, Agent, Hub, Hailuo, Audio, and the API, use How to Use MiniMax AI.
Context window and maximum output
The API catalog lists a 1,000,000-token context window for both M3 and M3.1 Flash. M2.7 was built for 204,800 tokens: that is the context length in its public model configuration and the window MiniMax listed for both M2.7 variants. The window is one budget for input and output together, not an input allowance followed by a separate full-size answer.[3][8]
system instructions + conversation history + tool definitions + tool calls and results + attached text / image / video tokens + model reasoning + final answer ≤ the model's context window
MiniMax’s Anthropic-compatible reference recommends max_tokens: 131072 for both M3 and M3.1 Flash and permits a hard setting of 524,288. The 65,536 recommendation and 204,800 maximum it gives for other models are the limits M2.x integrations worked with. These are output settings, not extra tokens beyond the context window, and thinking tokens count toward them. A 900K-token M3 input cannot also return a 524K-token completion.[4]
Five times the context does not mean five times the useful evidence
The nominal window of M3 and M3.1 Flash is about 4.88 times M2.7’s. It can fit workloads M2.7 could not accept in one request, but a larger ceiling does not guarantee perfect retrieval or reasoning across every token. Duplicate files, generated assets, irrelevant logs, and stale documentation can add cost and distract the model.
- Remove binaries, build output, vendored dependencies, and duplicate documents.
- Label files and evidence boundaries clearly.
- Ask for file, section, line, or quoted-evidence references.
- Test retrieval from the beginning, middle, and end of a long prompt.
- Compare full-context prompting with retrieval and staged summarization.
The larger window matters for jobs that need hundreds of thousands of source tokens. On M3, input above 512K tokens moves the call into a higher price band; M3.1 Flash has one flat rate.
Multimodality and reasoning controls
M3 and M3.1 Flash accept text, image, and video input and produce text. M2.7 was text-only through MiniMax’s OpenAI- and Anthropic-compatible APIs; its compatible message schema covered text, tools, tool results, and thinking content but not image or video blocks. For a text-only M2.7 integration, multimodal input is a new option, not a migration requirement.[4]
| Task | Better fit | Reason |
|---|---|---|
| Review a text-only pull request and test logs | Evaluate M3.1 Flash and M3 | Either can take over a text-only M2.7 task; compare correctness, retries, latency, and spend. |
| Compare a web implementation with a screenshot | M3.1 Flash or M3 | Both accept image input in the same model request. |
| Analyze a screen-recorded failure path | M3.1 Flash or M3 | Both accept video input; M2.7’s compatible text interface did not. |
| Low-latency classification with reasoning disabled | M3 | M3 exposes a documented no-thinking mode. M3.1 Flash always thinks (lower its effort instead), and M2.x thinking could not be disabled. |
| Keep an existing M2.7 agent’s behavior | Migrate and test, or self-host the M2.7 weights | Pay-as-you-go requests that name M2.7 fail after November 18, 2026, and plan requests go to M3.1 Flash. Any model switch can change tool selection, verbosity, reasoning, and parsing. |
Set M3 reasoning behavior explicitly. MiniMax documents different omission defaults across compatible endpoints: Anthropic Messages disables M3 thinking when the field is omitted, OpenAI Chat Completions uses adaptive thinking by default, and the Responses API disables reasoning when its reasoning object is omitted. Relying on omission can produce accidental latency and cost differences during migration.[12][13] M3.1 Flash has no off switch, so set its depth instead: output_config.effort on the Anthropic-compatible API, reasoning_effort on the OpenAI-compatible API, or reasoning.effort on the Responses API. Omitting it means max.[3]
In multi-step tool use, preserve the complete assistant response in history, including thinking or reasoning fields and tool-call blocks. Validate every tool argument, restrict permissions, cap steps and spend, and require approval before destructive actions. Model capability is not authorization to execute a command.[15]
Architecture, model size, and public weights
M3 and M2.7 both use a Mixture-of-Experts design, but M3 is not a simple weight update to M2.7. The official M3 card reports approximately 428 billion total parameters and about 23 billion activated per token. The M2.7 card reports 229 billion total parameters; its public configuration identifies 256 routed experts with eight selected per token.[7][8]
Full attention in M2.7 versus MSA in M3
M2.7 belongs to MiniMax’s full-attention M2 architecture line. M3 introduces MiniMax Sparse Attention (MSA), a block-sparse mechanism built on Grouped Query Attention. A lightweight index branch selects relevant key-value blocks for each attention group, and the main branch performs attention over those selected blocks rather than every preceding token.[14]
The MSA paper reports large compute and wall-clock improvements at one-million-token sequences, but its controlled architecture experiments use a 109B-parameter research model. Those results support the attention design; they are not an independent guarantee for every 428B M3 deployment, GPU, quantization, batch size, or API provider.[9]
Open-weight does not mean laptop-sized
MiniMax publishes both checkpoints, and the model cards list deployment routes including vLLM, SGLang, and Transformers. That makes private hosting and inspection possible, subject to each license. It does not make either model a routine laptop deployment. Sparse activation reduces per-token computation, but the full expert weights still need storage and a serving strategy. M3 has almost twice M2.7’s total parameter count and adds multimodal components.
Hosted and self-hosted results can diverge because of quantization, precision, kernels, context length, chat templates, tool parsers, sampling, parallelism, and vision/video support. Treat an engine or quantization change as a model-system change and rerun the evaluation suite.
If M3 is your self-hosting candidate, use the MiniMax M3 local deployment guide to size BF16 and quantized checkpoints, RAM, VRAM, long-context cache, vLLM, and SGLang before selecting hardware.
What the official benchmark claims do—and do not—show
MiniMax reports 59.0% for M3 on SWE-Bench Pro and 56.22% for M2.7 on SWE-Pro. The official Hugging Face evaluation metadata also displays approximately 59.0 and 56.2 for SWE-Bench Pro. That is a useful directional signal, but it is still provider-published evidence rather than a controlled evaluation performed by this website.[1][2][7][8]
| Benchmark | M3 result reported by MiniMax | M2.7 result reported by MiniMax | Comparison warning |
|---|---|---|---|
| SWE-Bench Pro / SWE-Pro | 59.0% | 56.22% | The roughly 2.78-point difference is vendor-reported; verify scaffold and run settings before treating it as an A/B result. |
| Terminal-Bench | 66.0% on Terminal-Bench 2.1 | 57.0% on Terminal Bench 2 | The benchmark versions differ, so the raw values are not directly interchangeable. |
| BrowseComp | 83.5 on the official M3 product page[16] | No directly paired M2.7 value on the cited product page | Do not invent a head-to-head delta where the same table does not provide one. |
| Long-running demonstrations | MiniMax reports multi-hour paper reproduction, model-training, and CUDA optimization tasks | MiniMax reports scaffold self-optimization and agent/productivity evaluations | These are first-party case studies, not independent replications. |
Benchmark outcomes can change with the agent scaffold, tool access, repository image, timeout, retry policy, pass count, judge model, prompt, and compute budget. The correct purchasing question is not which model has the larger headline score. It is which model completes your tasks correctly at an acceptable cost and with the least human correction.
MiniMax M3 vs M2.7 API price, with M3.1 Flash
The table below shows MiniMax’s standard pay-as-you-go rates for M3.1 Flash and M3, and the last rates MiniMax published for M2.7, taken from its pricing page as archived by the Wayback Machine on October 14, 2026. Rates are in U.S. dollars per one million tokens. MiniMax’s page labels the displayed M3 rates as a permanent 50% reduction from struck-through list rates; the table uses the billed amounts.[5][18]
| Hosted model / input band | Input | Output | Prompt-cache read | Prompt-cache write |
|---|---|---|---|---|
MiniMax-M3.1-Flash, any input length | $0.15/M | $0.60/M | $0.015/M | Not listed |
MiniMax-M3, ≤512K input tokens | $0.30/M | $1.20/M | $0.06/M | Not listed |
MiniMax-M3, >512K input tokens | $0.60/M | $2.40/M | $0.12/M | Not listed |
MiniMax-M2.7 (last published rate) | $0.30/M | $1.20/M | $0.06/M | $0.375/M |
MiniMax-M2.7-highspeed (last published rate) | $0.60/M | $2.40/M | $0.06/M | $0.375/M |
Key pricing conclusion: M3.1 Flash costs half of M2.7’s last standard rates on input and output ($0.15 versus $0.30, $0.60 versus $1.20) and 75% less than M2.7-highspeed’s ($0.15 versus $0.60 input, $0.60 versus $2.40 output), with cache reads at $0.015 instead of $0.06. M3 at or below 512K input matches M2.7’s last rates exactly; above 512K input, all three listed M3 rates double. Neither current model lists a cache-write price.
Two transparent cost examples
Assume standard service, no cache hits, no retries, and text token billing only.
Example A — 180K input + 10K output M3.1 Flash: 0.18 × $0.15 + 0.01 × $0.60 = $0.033 M3 (≤512K band): 0.18 × $0.30 + 0.01 × $1.20 = $0.066 M2.7, last rates: 0.18 × $0.30 + 0.01 × $1.20 = $0.066 M2.7-highspeed, last: 0.18 × $0.60 + 0.01 × $2.40 = $0.132 Example B — 600K input + 20K output M3.1 Flash: 0.60 × $0.15 + 0.02 × $0.60 = $0.102 M3 (>512K band): 0.60 × $0.60 + 0.02 × $2.40 = $0.408 M2.7: over its 204,800-token context window
These are calculations, not observed invoices. Images, video, reasoning, tool history, cache behavior, failed calls, service tier, and provider markup can change the total. On M3.1 Flash, thinking always runs and defaults to max effort; MiniMax’s advice for cutting its latency and token use is to lower the effort level.[3] Compare cost per successful task, not just the advertised rate per million tokens. See the broader MiniMax pricing guide for pay-as-you-go keys, Subscription Keys, Credits, and plan rules.
M Plan and Token Plan routing
Plan requests that name M2.5 or M2.7 have been answered by M3.1 Flash since October 19, 2026, under the old model names and with the same usage allowances, so plan users need no code change but should re-test the outputs they depend on. All three M Plan tiers include M3.1 Flash, and an M Plan API key (Subscription Key) is separate from a standard pay-as-you-go API Key; the two cannot be used interchangeably.[6][17] Our M Plan guide covers tiers, usage windows, and Credits.
Do not convert a per-token pay-as-you-go rate into a guaranteed M Plan or Token Plan allowance. Plan tiers, usage windows, shared modalities, and concurrency are separate rules. Check the billing console attached to the exact account before capacity planning.
M3 vs M2.7 license: the decisive self-hosting difference
Both models have downloadable weights, but neither should be described as unrestricted, permissively licensed open source. “Open-weight under a custom license” is the more accurate wording. Their commercial paths are different enough to affect model selection.
| License point | MiniMax M3 | MiniMax M2.7 |
|---|---|---|
| License name | MiniMax Community License | Non-Commercial License |
| Non-commercial use | Permitted subject to the license and retained notices | Permitted for personal use and non-profit research, subject to the license and prohibited-use policy |
| Commercial attribution | Prominently display “Built with MiniMax M3” | Prominently display “Built with MiniMax M2.7” if commercially authorized |
| Commercial use at or below the stated revenue threshold | The license requires a one-time notice to MiniMax; it states the threshold in relation to products or services generating no more than $20M in yearly revenue | Prior written authorization is required; the license provides no small-business revenue exception |
| Commercial use above the M3 threshold | Prior written MiniMax authorization is required | Prior written MiniMax authorization is required for commercial use at any size |
| Prohibited uses | Yes | Yes |
The M3 terms therefore provide a defined notice-based route for some commercial deployments that M2.7’s weight license does not. That does not make M3 equivalent to Apache 2.0 or MIT, and it does not eliminate attribution, notification, revenue, prohibited-use, or legal-review obligations.[10][11]
Keep the legal surfaces separate. Downloading, modifying, or serving the public checkpoint invokes the model-weight license. Calling MiniMax’s hosted API is governed by the applicable platform and paid-service terms. Paying API fees does not grant permission to self-host the weights commercially, and a weight-license summary does not replace the hosted service agreement.
This section summarizes source text and is not legal advice. Keep a dated copy of the exact license evaluated, check the checkpoint and derivative license, and obtain qualified legal review before commercial deployment or redistribution.
Decision matrix: which model should you choose?
| Your requirement | Start with | Why |
|---|---|---|
| New coding-agent integration | M3.1 Flash | Lowest listed rates ($0.15 / $0.60 per million tokens), a 1,000,000-token window at one flat rate, image and video input, and tunable effort |
| Pay-as-you-go system validated on M2.7 | M3.1 Flash or M3, after a regression run | Requests that name M2.7 fail after November 18, 2026, and a model switch can change output shape, tool behavior, reasoning, and latency |
| M Plan or Token Plan traffic that names M2.7 | No code change; re-test outputs | MiniMax has routed these requests to M3.1 Flash since October 19, 2026, with plan usage allowances unchanged |
| Workload that used M2.7-highspeed | M3.1 Flash, or M3 with Priority | MiniMax lists no highspeed ID for M3 or M3.1 Flash; M3.1 Flash is 75% below highspeed’s last published rates, and M3 Priority costs 1.5 times Standard for priority admission |
| Repository or document package above 204,800 combined tokens | M3.1 Flash or M3 | Both have 1,000,000-token windows; M3.1 Flash keeps one flat rate, while M3’s rates double above 512K input |
| Screenshot, diagram, or video understanding | M3.1 Flash or M3 | Both accept these input types through the documented compatible APIs |
| Calls that must skip thinking | M3 | M3 can disable thinking; M3.1 Flash always thinks, as M2.7 did |
| Self-hosted commercial product below M3’s stated $20M product/service threshold | M3 may have the clearer path | M3 specifies attribution plus notice; M2.7 requires prior written authorization for any commercial use |
| Permissive OSI-style license required by policy | Neither M3 nor M2.7 without legal approval | Both use custom licenses with commercial conditions and prohibited uses |
| Typical consumer laptop deployment | Neither the M3 nor the M2.7 checkpoint | Both are very large MoE models; M3 is substantially larger and multimodal |
| Exact reproduction of an M2.7 result | Self-hosted M2.7 weights (closest available rerun) | Pay-as-you-go requests that name M2.7 fail after November 18, 2026; hosted and self-hosted results can differ, so record the serving setup with the result |
For a new system, start with M3.1 Flash and keep M3 as the comparison when you need its Priority tier or thinking you can switch off. For an established M2.7 workflow on pay-as-you-go, run the replacement behind the same interface and switch once it passes your regression, safety, latency, and cost gates.
Migration checklist: M2.7 to M3 or M3.1 Flash
The base URLs do not change: M3.1 Flash and M3 use the same Anthropic-compatible (https://api.minimax.io/anthropic) and OpenAI-compatible (https://api.minimax.io/v1) routes that M2.7 used. On pay-as-you-go, the model ID is the change you cannot skip, because requests that name M2.7 fail after November 18, 2026. A production migration is still larger than a string replacement. See the MiniMax API guide for endpoint setup and SDK patterns.
1. Make the model configurable
# Replacement for the M2.7 model ID MINIMAX_MODEL=MiniMax-M3.1-Flash # Comparison candidate: swap in for the A/B run # MINIMAX_MODEL=MiniMax-M3
Keep the other model as a fallback while validation is in progress, and store each model’s settings (M3.1 Flash’s effort level, M3’s thinking mode) next to its ID so a switch changes both. Do not scatter a model ID through application code, tests, job definitions, and agent configuration.
2. Set thinking and effort explicitly
M3.1 Flash always thinks, as M2.7 did, and adds a depth setting: low, medium, high, xhigh, or max. Omitting it means max. Sending thinking: {"type": "disabled"} or effort none returns a 400 error.[3] On the Anthropic-compatible route, the request body looks like this:
{
"model": "MiniMax-M3.1-Flash",
"max_tokens": 8192,
"output_config": {"effort": "high"},
"messages": [{"role": "user", "content": "Review this diff for regressions: ..."}]
}On the OpenAI-compatible route, send reasoning_effort instead. On M3, choose adaptive or disabled thinking per task and API format: M3 has no effort control, and its thinking default is off in Anthropic Messages but on in OpenAI Chat Completions. Test both difficult reasoning work and simple low-latency calls.[4][12]
3. Update message and output parsers
- Preserve complete thinking, text, tool-use, and tool-result blocks.
- If your parser looked for
<think>tags incontent, update it. On the OpenAI-compatible route, M3.1 Flash returns thinking inreasoning_content,contentholds only the final answer, andreasoning_split=falseis not supported. - Confirm that structured output and tool arguments still pass schema validation.
- Add image/video handling only after validating supported content types, upload routes, and size limits.
- Cap output below both the configured maximum and the remaining combined context.
4. Recalculate context and cost guards
Raise the application context limit only when a task needs it. On M3, add an alert before 512K input because crossing that threshold changes all three listed standard rates for the call; M3.1 Flash has one flat rate, so its guard is total tokens and effort level. Keep retrieval, deduplication, and staged processing available rather than treating 1M as a target prompt size.
5. Run a controlled A/B evaluation
- Compare against recorded M2.7 results or a self-hosted M2.7 checkpoint.
- Use the same repository snapshot, prompts, tools, permissions, sandbox, timeout, and retry policy.
- Include bug repair, cross-file refactoring, test generation, repository navigation, ambiguous requirements, and tool failures.
- Require tests or deterministic checks to pass; do not score code by appearance alone.
- Record task success, invalid tool calls, regressions, input/output/cache tokens, time to first token, total wall time, retries, and human correction time.
- Calculate cost per accepted task, not cost per response.
6. Roll out gradually
Start with low-risk traffic, observe tool and parser failures, and keep the other model (M3 or M3.1 Flash) configured as the rollback. Revalidate privacy, rate limits, budget alarms, and authorization gates. A successful benchmark run is not a production approval.
A practical scorecard for your own comparison
| Metric | How to measure | Why it matters |
|---|---|---|
| Task success | Automated tests plus reviewer rubric | Primary evidence that the model solved the job |
| Regression rate | Previously passing tests broken by the patch | Measures collateral damage |
| Tool reliability | Valid calls / total calls; recovery from failures | Agent quality depends on execution, not prose |
| Human correction | Minutes and edits required before acceptance | A slightly slower model may reduce total labor |
| End-to-end latency | Task start to accepted result | Includes tools, retries, and reasoning |
| Total cost | Input + output + cache + retry + tool/provider costs | Per-token price alone can mislead |
| Long-context retrieval | Known evidence placed at multiple prompt positions | Tests whether extra context is usable |
| Safety | Unauthorized action attempts, secret exposure, unsafe commands | Strong coding ability increases the importance of controls |
Run several trials per task because sampling and tool environments introduce variance. Publish your scaffold, versions, prompts, pass count, and scoring rules if you later describe the result as an original benchmark.
Frequently asked questions
Is MiniMax M3 better than M2.7?
M3 has the stronger feature set: a 1M combined context window, image and video input, controllable reasoning, a newer sparse-attention architecture, and higher MiniMax-reported coding scores. For hosted use, the retirement settles the question: MiniMax closed M2.7 to new users on October 19, 2026, and API requests to it no longer work after November 18, 2026. The practical choice is between M3 and M3.1 Flash, and “better” should be decided by task success, correction time, latency, safety, and total cost on the same workload.
Does M3 cost more than M2.7?
Not at the standard pay-as-you-go tier while M3 input is at or below 512K tokens. M3 lists $0.30 per million input tokens, $1.20 per million output tokens, and $0.06 per million cache-read tokens, the same as the last rates MiniMax published for M2.7. Above 512K M3 input, the M3 rates double. M3.1 Flash is cheaper still: $0.15 input, $0.60 output, and $0.015 cache read per million tokens.
Is MiniMax M2.7 deprecated or retired?
Yes, on MiniMax’s hosted API. MiniMax closed the M2.5 and M2.7 families to new users on October 19, 2026, and set November 18, 2026 as their retirement date; after that date, API requests to MiniMax-M2.7 and MiniMax-M2.7-highspeed no longer work. On M Plan and Token Plan, MiniMax has routed M2.7 requests to M3.1 Flash since October 19, 2026, and the old names keep working. The open weights on Hugging Face are a separate route that the notice does not cover.
Can M2.7 accept images or video?
No. M2.7 is a text model: through MiniMax’s compatible APIs it accepted text and tool-call content, not image or video blocks. M3 and M3.1 Flash both accept text, image, and video input, and all three produce text output.
Does M3 provide five times more context?
M3 and M3.1 Flash both have a 1,000,000-token ceiling, about 4.88 times M2.7’s 204,800 tokens. The total includes input and output. Capacity does not guarantee perfect retrieval, and M3’s standard API rates increase when input exceeds 512K tokens, while M3.1 Flash keeps one flat rate.
Can I switch by changing only the model ID?
For a basic text call on the same compatible endpoint, replacing MiniMax-M2.7 with MiniMax-M3.1-Flash or MiniMax-M3 may be enough, and on pay-as-you-go the change is required, because requests that name M2.7 fail after November 18, 2026. Production migration also requires explicit reasoning or effort settings, parser checks, context and cost guards, tool-loop regression tests, and staged rollout. On M Plan and Token Plan, the old name keeps working because MiniMax routes it to M3.1 Flash.
Are MiniMax M3 and M2.7 open source?
The precise description is open-weight under custom licenses. M3 uses the MiniMax Community License; M2.7 uses a Non-Commercial License. Both allow downloading and permitted self-hosting, but neither is an unrestricted MIT- or Apache-style model-weight release. MiniMax’s retirement notice covers its hosted API and does not mention these weights.
Which license is easier for commercial self-hosting?
M3 provides a conditional notice-based route when the relevant products or services do not exceed the license’s $20M yearly-revenue threshold, with attribution required. M2.7 requires prior written authorization for any commercial use of the weights. Review the full license and obtain legal advice; hosted API terms are a separate contract.
Official sources used
- MiniMax M3 release report — release date, product positioning, MSA claims, multimodality, and MiniMax-reported benchmarks.
- MiniMax M2.7 release report — release date, self-improvement framing, coding, agent, and productivity positioning.
- MiniMax model invocation guide — hosted model IDs, context windows, compatible endpoints, and supported-model status.
- MiniMax Anthropic-compatible Messages reference — input types, output limits, thinking behavior, and content blocks.
- MiniMax pricing overview — M3.1 Flash and M3 pay-as-you-go rates, M3 input bands, and the Priority tier.
- MiniMax M Plan FAQ — M3.1 Flash in every M Plan tier, M Plan API keys versus pay-as-you-go API Keys, and Token Plan subscribers.
- Official MiniMax M3 model card — parameter counts, modalities, reasoning modes, weights, and deployment routes.
- Official MiniMax M2.7 model card — parameter count, weights, model configuration, evaluation metadata, and deployment routes.
- MiniMax Sparse Attention paper — MSA design and the scope of its 109B experimental results.
- MiniMax M3 Community License — commercial notice, attribution, revenue threshold, authorization, and prohibited uses.
- MiniMax M2.7 Non-Commercial License — commercial authorization, attribution, permitted free uses, and prohibited uses.
- MiniMax OpenAI-compatible Chat Completions reference — thinking controls for M3 and M3.1 Flash, including
reasoning_effort, and the M2.x thinking behavior used as the migration baseline. - MiniMax Responses API reference — M3 reasoning control and the no-reasoning omission default.
- MiniMax M2 full-attention engineering note — the M2 family’s full-attention design choice and limitations MiniMax identified in earlier efficient-attention experiments.
- MiniMax M3 tool-use and interleaved-thinking guide — preserving complete reasoning and tool-call content across turns.
- Official MiniMax M3 product page — product capabilities and MiniMax-reported BrowseComp result.
- MiniMax Model List — the M2.5 and M2.7 retirement notice (closed to new users October 19, 2026; retirement date November 18, 2026; M Plan and Token Plan routing to M3.1 Flash) and the Available and Retired Models tables.
- MiniMax pay-as-you-go pricing, Wayback Machine copy of October 14, 2026 — the last M2.7 and M2.7-highspeed rates MiniMax published.
- Deploy MiniMax-M2.7 — MiniMax’s self-hosting guide for existing M2.7 workflows.
- MiniMax M3.1 Flash release post — release date and product positioning.
Scope: This is a source-based comparison, not a hands-on benchmark or legal opinion. Model status, prices, context windows, API parameters, and code examples were checked against the linked official sources on November 9, 2026; benchmark, architecture, and license details were last checked on July 18, 2026.