MiniMax M2.5 is Here, and it's Free in Kilo for a Week
The new model is remarkably fast and excels at agentic engineering
We started Kilo Code with a simple mission: give every developer the world’s most powerful tools without the “frontier tax.” Today, we’re taking our biggest leap yet.
We are thrilled to announce that MiniMax’s hotly anticipated new model, MiniMax M2.5 is live on Kilo Code.
Not only is this a breathtaking new model from a top international lab, but we are making it completely free for all Kilo users for the first week.
Remember when we made M2.1 free for Kilo for Slack?
This is bigger.
Much bigger.
A Powerhouse That Can’t Be Ignored
MiniMax M2.5 isn’t just an incremental update; it is a total overhaul of the M2.1 architecture. While M2.1 has been the most popular open-weight model on Kilo to date, M2.5 moves MiniMax into the big leagues as a truly SOTA lab with a truly SOTA model. The OSS vs Proprietary distinction is blurring.
Scoring 51.3% on Multi-SWE-Bench, and 76.3% on BrowseComp, M2.5 is also more token efficient than previous generations, having been trained to optimize its actions and output through planning.
But the standout metric is really M2.5’s 80.2% performance on SWE-Bench Verified, a human-validated subset of real-world GitHub issues designed to test if AI can actually solve production-level bugs.
To put that in perspective, Claude Opus 4.6 averaged just below 80% across standard trials. While Anthropic noted that a specific prompt modification could push Opus 4.6 to 81.42% (and this fits with what we’re seeing with Opus in the wild), the fact that M2.5 sits comfortably at 80.2% out-of-the-box makes it a formidable powerhouse that rivals the best in the industry.
Beyond the flagship “Verified” test, M2.5 has proven its dominance in even more rigorous environments:
SWE-Bench Pro: M2.5 outpaced Gemini 3 Pro (55.4% vs 43.3%), proving it can handle the increased difficulty and realism of advanced software engineering tasks.
Multi-SWE-Bench: In complex, multi-step software suites, M2.5 again beat Gemini 3 Pro (51.3% vs 50.3%), showcasing superior autonomous execution in long-horizon tasks.
M2.5 Changes the Game — for the Better
MiniMax M2.5 was engineered from the ground up for the Agent-Verse. It is designed to be the primary workhorse for the future workspace, and at Kilo we are particularly excited about its agentic capabilities for both planning and executing large-scale dev projects.
Optimized for Agents: Specifically designed for agentic workflows, catching up to and often surpassing major models like GPT-5.2 and Gemini 3 Pro in coding and research.
Lightning Fast Throughput: Optimized for “thinking efficiency,” M2.5 clocks in at 100 tokens per second (tps)—achieving speeds 3x faster than Opus in early testing.
Always-On Efficiency: At $0.3/M input and a $0.06/M blended cost with cache, it offers the best price of any SOTA model for always-on agents. Of course, you’re getting it free in Kilo this week, but still good to know :)
Super Small: While other Tier-1 models require massive clusters, M2.5 utilizes only 10B activated parameters. It is the smallest Tier-1 model in existence, offering an unparalleled advantage for developers who want to self-host.
Try it Now on Kilo Code
We believe the best way to move the industry forward is to put the best models in the hands of every developer. Starting today, you can select MiniMax M2.5 in the model dropdown and start building immediately. No credits required—just pure, unadulterated SOTA power.
For the past month, MiniMax M2.1 has been the top model on Kilo’s leaderboard for every mode except for Architect and Orchestrator. Will M2.5 push MiniMax over the edge to rule every category?
Try M2.5 in Kilo for free today.