MiniMax Review (2026): Features, Pricing & Verdict

What Is MiniMax?

MiniMax is the Shanghai AI company, founded in 2021 and one of the original “AI tiger” startups of China’s post-ChatGPT boom, that took the road nobody else in this category took: instead of betting everything on one text model, it built a consumer-proven engine for every modality and then opened the whole stack to developers. Most Western users met MiniMax without the name — through Hailuo, the video generator whose speed-and-physics act we scored highly at 0379; through its speech models, widely regarded as some of the best text-to-speech in the world and the default narrator of a substantial share of AI-generated content; or through its companion apps (Talkie internationally, Xingye in China) that proved the company’s conversational chops at consumer scale. The text line — the M-series — is what earns MiniMax its slot in this category, and its story is an efficiency thesis executed consistently. M2 (23 October 2025) set the template: a 230-billion-parameter Mixture-of-Experts model activating just 10 billion per token, optimised end-to-end for coding and agentic work — multi-file editing, compile-run-fix loops, test-validated repair — with strong SWE-Bench Verified, Multi-SWE-Bench and Terminal-Bench results, competitive agentic scores on BrowseComp and GAIA, and a top-open-source ranking for composite intelligence from Artificial Analysis, all at $0.255–0.30 input/$1.00–1.20 output per million tokens with a ~205K context. The line then iterated fast and stayed cheap: M2 Her (January 2026) spun the conversational-EQ heritage into a dedicated companion model; M2.7 (18 March 2026) became the workhorse — same 230B/10B skeleton, sharper agentic training with multi-agent collaboration, and a benchmark profile aimed squarely at real-world productivity: 56.2% on SWE-Pro, 57.0% on Terminal-Bench 2, and a 1495 ELO on GDPval-AA, the evaluation built around actual office deliverables (live debugging, root-cause analysis, financial modelling, full document generation across Word, Excel and PowerPoint) where M2.7 set the standard for multi-agent systems; and M3, the current flagship, extends to frontier coding-and-agent positioning with an advertised 1M-token window (tiered above 512K) at the same effective $0.30/$1.20 under a labelled-permanent 50% discount. Distribution is broad — an OpenAI-compatible first-party API, a dozen-plus providers on OpenRouter (where M2.7 lists at $0.24/$0.96 and spot markets have driven it lower still), and weights on Hugging Face, though the licence tightened from M2’s true MIT to a modified MIT barring commercial self-hosting from M2.5 onward. Around the text line sit the other meters: speech, video, music and image APIs, each billed on its own terms — making MiniMax the only provider in this category where “add narration and a video clip to the agent’s output” is a same-platform decision. Within our Model Providers & AI Infrastructure category, MiniMax is the multimodal value house: not the price floor (DeepSeek), not the breadth king (Qwen), not the swarm specialist (Moonshot), but the one lab whose answer to text, voice and video is the same invoice.

Core Features

The M-series: office-work agents at 5% of flagship cost

The M-series’ pitch is precision-targeted: it isn’t trying to win abstract reasoning crowns — it’s built to do jobs, and the benchmark that defines it is the one closest to actual employment. GDPval-AA evaluates models on real-world knowledge-work deliverables — debugging live systems, root-cause analysis, financial modelling, generating complete Word documents, Excel workbooks and PowerPoint decks — and M2.7’s 1495 ELO set the standard for multi-agent systems on it, which is a more commercially meaningful claim than another decimal point on a math olympiad. The supporting numbers hold the frame: 56.2% on SWE-Pro (the harder autonomous-engineering benchmark where even flagships live in the 50s–60s), 57.0% on Terminal-Bench 2, an 87.4 GPQA score placing it in the 94th percentile, and M2’s Artificial Analysis composite ranking among the top open-source models — with the agentic fundamentals (long-horizon planning, retrieval, recovery from execution errors on BrowseComp and GAIA) trained in from the start rather than prompted on top. M2.7’s multi-agent collaboration layer echoes Moonshot’s swarm thesis at smaller scale: the model plans, executes and refines across cooperating instances, which is precisely what the document-generation workflows exercise. Now the economics that make it interesting rather than merely good: the 230B-total/10B-active architecture means inference costs like a 10B model, so the entire M-series bills $0.30 input/$1.20 output per million tokens — about 5% of Opus-class output rates, cheaper on output than Kimi K2.6 ($2.50) and undercut on input only by DeepSeek’s V4 Flash — and the version number changes capability, never price. An agent generating a hundred client reports a day, or a support system producing full spreadsheets on demand, runs at costs that would be a rounding error on a Western flagship. The honest capability boundaries: this is frontier-adjacent, not frontier — the hardest single-shot reasoning still belongs to the closed Western models and DeepSeek’s V4 Pro; the ~205K context (M2/M2.7) trails DeepSeek’s and Qwen’s million-token windows, and M3’s advertised 1M is tiered (double price above 512K) and gate-limited at the top band; standard output speed of ~48 tokens/second is the slowest in the Chinese cohort — fine for background agents, awkward for user-facing streaming without the 2x HighSpeed variant; and the interleaved-thinking design imposes a real integration requirement: reasoning content must be passed back between turns (via reasoning_details on routers) or multi-turn performance degrades, a quirk that has tripped more than one naive OpenAI-SDK port, alongside documented tool-schema edge cases. Build for what it is — a tireless, absurdly cheap office worker with a slight stutter — and the M-series is one of the best deals in this review.

The multimodal stack: one vendor, five meters

MiniMax’s structural advantage over every text-first rival in this category is that the text API is one aisle in a larger store — and for a growing class of products, that consolidation is the decision. The inventory: speech — MiniMax’s text-to-speech and voice-cloning models are widely regarded as among the best in the world, with naturalness and multilingual range that made them the de facto narrator of AI-generated video content well before most users knew the vendor’s name; video — Hailuo, reviewed separately at 0379, is the budget speed-and-physics champion of the generative-video market, and its API lives on this same platform; music and image generation round out the creative set, each billed on its own meter (per-character or per-second or per-image, not tokens — read each price page). Why this matters architecturally: the canonical 2026 product — an agent that researches, writes, narrates and illustrates — normally requires stitching three or four vendors (a text API, ElevenLabs-class TTS, a video model, an image model) with separate keys, invoices, rate limits and compliance reviews. On MiniMax it’s one platform, one account, and text-model outputs that were explicitly trained toward document and media production feeding sibling APIs. The consumer provenance is the quality assurance: these aren’t checkbox features — Hailuo and the companion apps (Talkie/Xingye) run at consumer scale, so the speech and video engines have survived contact with millions of users, and M2 Her (January 2026) productised the emotional-conversation heritage into a dedicated 66K-context companion model at the standard $0.30/$1.20 — a niche no other provider in this category serves as a first-class API. The platform machinery beneath: OpenAI-compatible endpoints, two parallel billing systems (pay-as-you-go on standard keys; Token Plan subscriptions on separate keys with fixed quotas managed in 5-hour rolling and weekly windows, unused quota not carried over), a Priority service tier on M3 (1.5x for admission-priority latency), and prompt caching with cheap reads ($0.06/M) but a write cost ($0.375/M) that exceeds standard input — so caching pays only on genuinely repeated prefixes. The honest limits of the breadth claim: the modality APIs are siblings, not a fused any-to-any model — you orchestrate calls across them rather than prompting one omnimodel; the consumer apps (MiniMax Agent, Hailuo) run on entirely separate credit subscriptions that confuse first-time buyers; and Qwen technically lists more model IDs — but Qwen’s media models are catalogue entries, while MiniMax’s voice and video are market leaders in their own reviews. For media-adjacent builders, that’s the distinction that decides.

Licensing, speed and trust: reading the fine print

MiniMax’s fine print carries more signal than most in this cohort, and three clauses deserve the spotlight. First, the licence regression — the most consequential and least advertised change: M2 shipped under a true MIT licence, full commercial self-hosting included, and became a legitimate open-weights citizen on that basis; from M2.5 onward (including M2.7), the weights moved to a modified MIT that permits free personal use but bars commercial deployment without separate authorisation from MiniMax. The practical fork: commercial self-hosters can stay on the older MIT-licensed M2/M2.5-era terms they already hold, route commercial traffic through the paid API (which is what the licence is designed to encourage), or negotiate authorisation — and any architecture that assumed “MiniMax weights are MIT” needs re-reading before the next deployment. Set against DeepSeek’s unqualified MIT and Moonshot’s lightly-modified MIT on their flagships, this is the weakest openness posture of the Chinese big four, and it moves MiniMax’s open-weights story from “portability guarantee” to “evaluation privilege.” Second, speed and its price: ~48 tokens/second standard output is a real product constraint — background agents won’t care, user-facing chat will — and the remedies are explicit line items (M2.7-HighSpeed at exactly 2x; M3 Priority at 1.5x), which means latency-sensitive workloads should be budgeted at the higher rates from day one rather than discovering the markup later. Third, the promo-price pattern: M3’s headline $0.30/$1.20 is documented as a “permanent 50% off” against a $0.60/$2.40 list — and this category’s recent history (Qwen’s expiring flagship promo, the April free-tier purges, versus DeepSeek’s discounts-become-list-price) argues for budgeting at list until “permanent” survives a few quarters. On residency and trust: MiniMax is a China-headquartered vendor, and the standard cohort analysis applies — organisations barring Chinese-parent hosted platforms will bar the first-party API; the mitigation is thinner than peers’ because the modified licence complicates the self-host route for commercial use, leaving Western aggregators and OpenRouter’s dozen-plus providers (M2.7 from $0.24/$0.96, with spot markets lower) as the main non-China inference path. The credibility column is genuine: consumer-scale battle-testing across Hailuo and the companion apps, a text line iterated four times in six months without a price increase, top-percentile independent benchmark placements, and OpenRouter traffic that confirms real production adoption. The posture to take: treat MiniMax as an outstanding API vendor with a leading multimodal stack — and unlike its peers, do not count the open weights as your exit strategy unless your use is personal or your lawyers have read the modified licence.

Scored Categories

Pricing

Frequently Asked Questions