MiniMax M2.5 in Cline: Multi-Agent Coding at SOTA Speed

TL;DR

• MiniMax released M2.5, now available in Cline across all platforms (VS Code, JetBrains, Zed, Neovim, Emacs, CLI)
• Built for multi-agent workflows with 100 tokens/sec throughput and $0.06/M blended cost — 3x faster than Opus at a fraction of the price
• Free for a limited time; select minimax-m2.5 from the model picker and add your MiniMax API key

What Dropped

MiniMax released M2.5, a 10B-parameter model trained specifically for multi-agent coordination and workspace fluency. It's live in Cline across VS Code, JetBrains, Zed, Neovim, Emacs, and the CLI. The headline: it beats Claude Opus 4.6 on SWE-Bench Pro (55.4 vs 53.4) while running 3x faster and costing a fraction as much.

The Dev Angle

M2.5 was trained to work alongside other agents, not assume it's the only one in the room. That matters for Cline Subagents workflows where you run multiple agents in parallel on different parts of a task. The model context-switches cleanly between writing code, reviewing tests, and generating docs without losing track of which task it's on.

The performance numbers: 100 tokens per second (roughly 3x faster than Opus), $0.30/M input tokens, $0.06/M blended with caching. On real-world benchmarks, M2.5 edges Opus on multi-repo coordination (51.3 vs 50.3 on Multi-SWE-Bench) and beats it decisively on complex engineering tasks (55.4 vs 53.4 on SWE-Bench Pro). Opus still leads on terminal tasks and multilingual coding, but the gap is narrow.

The 10B activated parameter footprint is the smallest at this tier by a wide margin. For teams running Cline Enterprise with VPC deployments, that means SOTA-level coding performance on hardware that would have been undersized for previous-generation models.

M2.5 also handles Excel, Word, and PowerPoint natively alongside code. If your workflow spans writing code and building analysis spreadsheets or drafting summary docs, you stay in one agent instead of context-switching between models.

Should You Care?

If you're running Cline subagents or the CLI with auto-approve: This is the move. The speed and cost combination makes always-on agents practical. You're not watching the meter anymore. You give it a task, it executes, feedback loop is tight. The model doesn't over-explain or second-guess — it plans, acts, moves on.

If you're using Cline in an IDE for interactive coding: M2.5 is worth a test, especially if you're on Opus. You'll see faster response times and lower costs. The multi-agent training won't matter as much in single-agent workflows, but the raw performance does.

If your work stays in the codebase: The workspace fluency (Excel, Word, PowerPoint support) won't change your workflow. The speed and cost will.

If you're self-hosting: The 10B footprint opens up deployment options that weren't viable before. Check with your infrastructure team about what hardware you can now run SOTA models on.

MiniMax is offering M2.5 free for a limited time. There's no reason not to try it. Select minimax-m2.5 from the model picker in Cline, add your MiniMax API key, and run it on a subagent workflow or CLI task to see the difference.

Source: Cline