Mavis

The current desktop product lets users create agents with separate roles and working directories, assemble them into teams and run them concurrently. MiniMax positions this architecture for long-running research, coding and professional work where parallelism and specialization matter. A verifier role can review other agents' work before delivery.

Mavis sits inside the broader MiniMax model stack rather than being a fixed-model product. Current plans expose MiniMax M3/M2.7 plus image, speech and music capabilities, a one-million-token context window and multimodal inputs. Users can also use the same subscription quota through MiniMax Code or compatible external tools/API paths. ¹

MiniMax adjusted Token Plan pricing on August 25, 2026. Current monthly individual tiers are Plus at $20, Max at $50 and Ultra at $120. They include progressively larger M3-equivalent token quotas and higher concurrent-agent limits: roughly 3-4 agents on Plus, 4-5 on Max and 6-7 on Ultra. Annual billing is discounted. Video generation is included only on higher tiers, while text, image, speech and music share quota. ²³⁴

Independent exact-product evidence is limited but not absent. KDnuggets published an August 2026 hands-on evaluation that correctly distinguishes the new Mavis/Agent Teams product from older MiniMax Agent coverage. It found the multi-agent architecture useful for decomposing complex work but emphasized that output quality still depends on task decomposition, verification and the underlying model rather than the orchestration layer alone.

A public MiniMax Code GitHub issue documents a reproducible Windows startup conflict affecting the bundled Mavis command after auto-update. One issue is not a reliability benchmark, but it is concrete product-level evidence that desktop integration can have rough edges.

Trustpilot currently shows MiniMax at about 1.8/5 from 26 reviews. The profile spans the broader MiniMax service rather than Mavis alone, so its score should not be read as an exact-Mavis product measure. It is still material service evidence: recent complaints include support non-response, billing/refund friction, outages and poor coding-agent reliability. The sample is small and self-selected, so it should be treated as a warning signal rather than a representative satisfaction measure.

Current independent exact-Mavis evidence remains limited. Trustpilot is a broader MiniMax service signal rather than an exact-product rating, while company/model coverage does not establish Mavis product performance. Model benchmarks for M2.7/M3 are not automatically transferred because orchestration, tools, context and verification materially change the product surface. ⁵⁶⁷⁸

Mavis is best suited to developers and technical teams comfortable with agentic workflows who want several agents working in parallel over local projects or long-running tasks. It is particularly attractive to existing MiniMax users because one plan now covers model/API, coding and agent usage.

It is less suitable for teams that require a mature independent review corpus, proven desktop reliability or enterprise-grade support evidence before deployment. High-stakes work should retain human review even when the verifier agent is enabled.