Fireworks AI unveils MiniMax M2.7

Fireworks AI has added MiniMax M2.7 to its model library, making the new model available now through serverless and dedicated deployments. (fireworks.ai) Fireworks lists M2.7 as a 228 billion-parameter mixture-of-experts model with a 196,600-token context window, function calling, image input support, and fine-tuning support. Fireworks prices it at $0.30 per 1 million input tokens, $0.06 per 1 million cached input tokens, and $1.20 per 1 million output tokens. (fireworks.ai) Large language models predict the next chunk of text; a mixture-of-experts model routes each request to a smaller set of specialized sub-models instead of using the full network every time. Fireworks and MiniMax pitch that design as a way to handle long, tool-heavy workflows without the full cost of a dense model of similar size. (fireworks.ai, minimax.io) MiniMax said on March 18 that M2.7 was built for “complex agent harnesses,” meaning software setups where a model uses tools, memory, and role-based sub-agents to complete multi-step jobs. The company said it used that approach internally to let the model update memory, build skills, and improve parts of its own training workflow. (minimax.io, github.com) That framing puts M2.7 in the middle of a crowded market for coding and agent models, where providers are competing on benchmark scores, token prices, and how quickly developers can plug a model into production systems. Fireworks’ own recommendation guide, updated this month, still lists MiniMax 2.5 among its suggested options for code, agents, and long-context work. (docs.fireworks.ai, fireworks.ai) MiniMax says M2.7’s strongest pitch is software engineering work that goes beyond autocomplete. In its GitHub repository, the company says the model handles log analysis, bug troubleshooting, refactoring, code security review, database root-cause checks, and site reliability engineering decisions. (github.com) The company also published benchmark claims aimed at buyers comparing it with closed models. MiniMax says M2.7 scored 56.22% on SWE-Pro, 57.0% on Terminal Bench 2, 55.6% on VIBE-Pro, 76.5 on SWE Multilingual, and 52.7 on Multi SWE Bench. (github.com, minimax.io) MiniMax ties those scores to a broader “self-evolution” pitch. In its repository, the company says an internal version of M2.7 optimized a programming scaffold over more than 100 rounds and improved performance by 30%, though that result is a company-reported internal test rather than an independent evaluation. (github.com) For Fireworks, the immediate move is straightforward: put a newly released open model in front of developers on day one, with a live endpoint, pricing, and deployment options already attached. For developers building coding copilots, security tools, or operations assistants, that means M2.7 is now one more model they can test against incumbents instead of waiting for a later integration. (fireworks.ai, fireworks.ai)