Table of Contents
🌏 中文版
MiniMax didn't start as a model company. It built character chat apps (Talkie) and video generation tools (Hailuo AI), earning its spot as one of China's "Six AI Tigers." Then in February 2026, its M2.5 scored 80.2% on SWE-bench Verified — just 0.6 points behind Claude Opus 4.6 at the time — at one-tenth to one-twentieth the API price. Here's how a chat app company ended up on the coding frontier.
Company Background
MiniMax was founded in December 2021 in Shanghai by Yan Junjie, formerly of SenseTime's computer vision team. The name comes from the Minimax algorithm — making optimal decisions by accounting for the opponent's best response.
As of 2025, the company has about 415 employees, revenue of roughly $79 million, and is still unprofitable (operating loss of $1.87 billion). Funding milestones: $600M led by Alibaba in March 2024 ($2.5B valuation), HKEX IPO in January 2026, and another $2B raise in July 2026.
Why would a consumer AI company build coding models? MiniMax's logic: to make AI products (chat, roleplay, video) truly good, the underlying model needs strong reasoning and tool-use capabilities. Coding ability is the foundation of agentic ability — models that write code well tend to use tools well.
Model Family
| Model | Released | Architecture | Total Params | Active/Token | Context | Highlight |
|---|---|---|---|---|---|---|
| M2.5 | 2026-02 | MoE | ~229B | Undisclosed | — | SWE-bench 80.2%, price/perf benchmark |
| M2.5-Lightning | 2026-02 | MoE | ~229B | Undisclosed | — | 2× throughput (100 tok/s) |
| M2.7 | 2026-03 | MoE | — | ~10B | — | Ultra-efficient, SWE-bench 78% |
| M3 | 2026-06 | MoE + MSA | 456B | 45.9B | 1M | Open-weight flagship |
M2.5: The Price/Performance Story
M2.5's story isn't about having the highest score — it's about how little it costs to get that score.
| Metric | M2.5 | Claude Opus 4.6 (contemporary) |
|---|---|---|
| SWE-bench Verified | 80.2% | 80.8% |
| API price (input / output) | $0.15 / $1.20 per M tokens | ~$15 / $75 per M tokens |
| Price gap | — | ~10-60× |
Per MiniMax's official claims, "$1 runs continuously for 1 hour at 100 tok/s." Four M2.5 instances running 24/7 for a year cost roughly $10,000.
During training, M2.5 exhibited an interesting emergent behavior: the model spontaneously writes architecture specs before coding (a "spec-writing tendency"). This wasn't engineered — it appeared naturally through extensive RL training.
M3: The Technical Leap
M3 is MiniMax's largest model and their technical flagship.
Per MiniMax's official technical report, M3's core innovation is MiniMax Sparse Attention (MSA): it replaces full attention computation with KV-block selection, reducing per-token compute at long contexts to roughly 1/20th of the previous generation. This makes a 1M context window economically feasible.
M3 is also a natively multimodal model, accepting text, image, and video inputs.
Benchmark Comparison
All data from official MiniMax releases.
M2.5 vs Contemporary Competitors
| Benchmark | M2.5 | Claude Opus 4.6 |
|---|---|---|
| SWE-bench Verified | 80.2% | 80.8% |
| Multi-SWE-Bench | 51.3% | — |
| BrowseComp (with context mgmt) | 76.3% | — |
M2.5 trailed Opus 4.6 by just 0.6 percentage points, at an order-of-magnitude lower cost.
M3 vs Open/Closed Flagships
| Benchmark | M3 | GPT-5.5 | Ornith 1.5-397B |
|---|---|---|---|
| SWE-bench Pro | 59.0% | 58.6% | — |
| GPQA Diamond | 92.9% | — | 92.8 |
| HLE | 39.0% | — | 44.6 |
M3 was the first open-weight model to surpass GPT-5.5 on SWE-bench Pro. Compared to fellow MoE model Ornith 1.5-397B, the two trade blows on GPQA Diamond but take entirely different technical paths: Ornith uses self-improvement RL, while MiniMax uses massive-scale environment RL.
M2.7: The Overlooked Efficiency Play
M2.7 deserves a mention: activating only ~10B parameters per token, it scores 78% on SWE-bench Verified — in the same range as Ornith 1.5-35B-A3B's 79%, though with higher activation (10B vs 3B). Both are competing on the same "use MoE to minimize inference cost" track.
Training Methodology
MiniMax's RL training takes a different path from Ornith's self-improvement loop.
Per MiniMax's official documentation, M2.5 was trained via RL across "over 200,000 complex real-world environments" spanning 10+ languages. These aren't synthetic exercises — they're environments extracted from real software projects where the model attempts to solve problems, with success or failure as the reward signal.
M3 introduces an interactive user-simulator framework: it trains collaborative abilities through multi-turn dialogue scenarios using simulated users. Unlike most coding models that train on single-turn problem solving, M3 also learns to interact, ask questions, and iterate on modifications.
Pricing Strategy
MiniMax's pricing is its sharpest differentiator.
| Model | Input | Output | Speed |
|---|---|---|---|
| M2.5 | $0.15/M | $1.20/M | Standard |
| M2.5-Lightning | $0.30/M | $2.40/M | 100 tok/s |
| M3 | $0.60/M | $2.40/M | — |
The gap versus closed-source flagships ranges from 10× to 60×. This pricing makes "running your coding agent on MiniMax" a serious option for many teams — especially for batch processing scenarios requiring high API call volumes.
Is It Worth Watching?
Yes, for three reasons:
- Cost structure proof-of-concept — M2.5 proved that frontier coding doesn't require frontier pricing. For dev teams, a 10× cost difference doesn't just save money — it shifts which use cases are economically viable
- MSA's engineering significance — 1M context isn't new, but achieving it at 1/20th the compute is. If MSA's quality trade-offs are as minimal as MiniMax claims, this architectural direction is worth tracking
- Consumer AI feeding back into foundation models — MiniMax's user base (200M+) provides training signals other pure model companies lack. The interactive user-simulator framework may be a direct expression of this advantage
Caveats:
- The company is still deeply unprofitable ($1.87B operating loss) — pricing may be subsidized
- M3 is open-weight but training data and code are not released — not fully open-source
- Benchmark scores are self-reported; large-scale independent reproductions are ongoing
- Geopolitical risks around Chinese AI companies are a factor for some use cases
References
Loading...