Every MiniMax AI Model Explained and Compared (Nov, 2026)

The MiniMax Lineup in Plain English

MiniMax names its families by letter. M is the language line, which reads text, images and video and writes text. H is the generation line: H3 makes video with sound, and H3 Image makes pictures. Speech and Music keep plain names.

The Hailuo brand, which MiniMax used for its video models through Hailuo 2.3, now sits in the legacy list, and H3 replaces it.

Most of the line is open weight. The MiniMaxAI organization on Hugging Face hosts M3, M3.1-Flash, H3, H3-Image and Music 3, next to older checkpoints back to MiniMax-Text-01 from January 2025.

The speech models and the hosted helpers, such as the prompt rewriter H3 leans on, stay API only.

MiniMax-M3.1-Flash

Flash in MiniMax's naming does not mean small. The M3.1-Flash model card lists about 428B total parameters with 23B active per token, the same size as M3.

The serving stack is the difference: sparse attention through the whole model, quantization-aware training and a speculative decoding head.

MiniMax says online inference runs nearly three times faster than its predecessor and agent tasks cost 70% less, with output peaking above 200 tokens a second.

It costs $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 for cached input, with one flat rate across the full 1M-token window.

MiniMax's own benchmark table puts it at 73.7% on DeepSWE v1.1 and 61.7 on OSWorld 2.0. The same table is candid about the gap: on Terminal Bench 4.0 it scores 38.1 against 51.8 for Claude Opus 5.

Those are vendor-run numbers, so test on your own tasks before you trust them.

MiniMax-M3

M3 arrived on June 1, 2026 and introduced MiniMax Sparse Attention, an operator built for million-token prompts.

According to the M3 model card, it runs 9 times faster at prefill and 15 times faster at decode than M2 at 1M context, cutting per-token compute to a twentieth.

MiniMax reports 80.5 on SWE-bench Verified and, on its M3 page, 83.5 on BrowseComp against 79.3 for Claude Opus 4.7.

M3 stays on the API for two reasons. It is the only current text model that can answer without thinking, and it is the only one with a Priority tier.

Its context promise is also hedged: the API supports up to 1M tokens with a guaranteed minimum of 512K, while MiniMax lists M3.1-Flash at a flat 1,000,000.

MiniMax-H3 video

H3 makes 4 to 15 second clips at 24 frames a second, with voice, sound effects and music generated in the same pass as the picture.

The H3 model card describes a 33B-parameter dense transformer that outputs 768P, plus a second pass that regenerates the clip at 2K. Not every part is open.

MiniMax keeps the prompt rewriter, H3-Context-IR, on its own servers, at $0.90 input and $3.60 output per million tokens, and has not released the 2K pass as weights.

Users in the US, EU, UK and South Korea fill in an application form before using the weights.

On the API, H3 bills by the second of video, not by the clip. Four variants trade quality for price, from H3 at 2K down to H3-Lite at 480P.

Per-second billing changes the math against the old Hailuo models, which are still on the API.

At $0.08 a second, a 10-second H3 clip at 768P comes to $0.80 of output, while the same pricing page lists a 10-second 768P Hailuo 2.3 clip at $0.56 flat. H3-Lite at 768P works out to $0.40 for those 10 seconds.

H3-Max costs the same $0.08 a second for output but bills reference video at $0.143 a second, and it is the variant that accepts image, video and audio references together.

MiniMax-H3-Image

H3-Image is the newest model in the family, released on November 10, 2026. It handles text to image and editing with 1 to 16 reference images, and the open checkpoint on Hugging Face is a 33.1B-parameter diffusion transformer.

On the API it bills by tokens. A 1K text-to-image render costs about $0.022 on Standard and $0.035 on Highspeed, by MiniMax's own worked examples. A 50% discount on 4K output runs until November 17, 2026.

The older image-01 still costs $0.0035 per image, which makes it the budget pick for thumbnails.

Speech 2.8 and the voice tools

Speech 2.8 launched on January 23, 2026 with sound tags that let a voice laugh, breathe or say "um" on cue, according to the Speech 2.8 announcement, which also says a 10-second sample is enough to clone a voice.

Both versions cover 40 languages. speech-2.8-hd costs $100 per million characters and speech-2.8-turbo costs $60. Cloning a voice is $1.50 per voice and designing one from a prompt is $3, each charged on first use.

Speech-to-text runs $0.38 per hour of audio.

Music 3

Music moved off the paid API in 2026. New users lost access to music and lyrics generation on August 20, 2026, and the model list now files every music model under retired, though existing paying customers can keep calling them.

The replacement is Music 3 as open weights: an 8B planning model built from Qwen3-8B, a 0.6B model for acoustic detail, and a synthesis stage that writes 32 kHz stereo WAV.

MiniMax Pricing Compared Across the Lineup

Output tokens drive a text bill, and M3.1-Flash charges half what M3 does for them: $0.60 against $1.20 per million. M3's listed rates are already half its original prices, under what MiniMax labels a permanent 50% discount.

Long prompts cost more on M3 only: past 512K input tokens, its rates double to $0.60 and $2.40. Priority adds 50% on top, and M3.1-Flash does not offer it.

Cached input is the other lever.

On M3.1-Flash a cache hit costs $0.015 per million tokens, a tenth of the input rate, and M3 charges $0.06 up to 512K. An agent that resends the same system prompt and file context on every turn pays mostly cache rates after the first call.

MiniMax does not list a separate cache write price.

Subscriptions sit beside pay-as-you-go. The Token Plan comes in Go, Explore and Build tiers with 1x, 3x and 7.5x the base quota. All three include M3.1 Flash, H3 Image and speech; only Explore and Build include H3 video.

Quotas reset on a 5-hour and a weekly window, and unused quota does not roll over. MiniMax does not print the tier prices in its docs, so check the console before you budget.

How Thinking and Effort Work

The two text models handle reasoning in opposite ways, per the text generation guide. M3.1-Flash always thinks.

You set depth with effort, at low, medium, high, xhigh or max, and leaving it out means max, the slowest and most token-hungry setting. Sending a disabled flag returns a 400 error.

On the Anthropic-style API the field is output_config.effort; on the OpenAI-style API it is reasoning_effort.

M3 starts with thinking off. Pass adaptive and the model decides when to reason; it has no effort dial. The practical rule: set effort on every M3.1-Flash call instead of trusting the default.

Start at medium for chat and extraction, and keep max for planning and hard code changes. If a route must answer with zero reasoning tokens, it belongs on M3.

Which MiniMax Model Should You Use?

If I were starting a new agent project this month, I would begin on M3.1-Flash with effort set explicitly, because it is the cheaper and newer of two models that share one parameter count. The table maps common jobs to a model.

Measure before you switch models. Lowering effort on M3.1-Flash is cheaper than moving traffic to M3, and moving M3 traffic to Priority costs 50% more on every token.

For a wider view of how Chinese labs price against US ones, see our US vs China AI model cost comparison, and for MiniMax next to its closest Chinese peers, read the Kimi model guide and the DeepSeek model guide.

What This Means for Hiring

MiniMax makes the model cheap; the engineering around it is where the cost sits. Routing between effort levels, migrating off M2.7 before the cutoff, and serving a 428B open model take engineers who have shipped LLM systems.

Our LLM developer cost guide puts a mid-level LLM developer at $119 to $185 an hour in the US, against $23 to $39 an hour for a vetted LLM developer from Asia. We send a shortlist within 24 hours.