MiniMax M3 Just Toppled GPT-5.5 and Released Open Weights
Looks like the folks at MiniMax decided to throw a party and invited everyone to watch GPT-5.5 get pushed off its pedestal. It’s a bold claim, a mountain of open data, and exactly the kind of chaos this industry needs to keep things interesting.
The newly unveiled M3 model relies on a proprietary architecture dubbed MSA (MiniMax Sparse Attention). This tweak allows the model to process a massive 1-million-token context window while slashing computational requirements by a factor of 20 compared to its predecessor. The efficiency gains are palpable, with prefill speeds jumping 9x and decoding speeds hitting 15x faster than before.
In the arena of raw performance, M3 clocked a 59.0% score on SWE-Bench Pro, effectively nudging past GPT-5.5 and Gemini 3.1 Pro in coding tasks. It also claimed the top spot in Claw-Eval for agentic workflows and maintained a 66.0% success rate on Terminal-Bench 2.1. While Gemini 2.5 Pro reportedly still edges it out in context length, M3 is arguably the most capable model currently offering open weights.
Beyond just crunching code, the model features native multimodal capabilities trained from scratch on a staggering 100 trillion tokens of text, images, and video. Through the MiniMax Code interface, the system can actively manipulate desktop environments, opening files and running commands with frightening precision. Developers can access this via API right now, with the full technical report and open weights slated to drop within the next 10 days.
The era of closed-garden dominance is predictably eroding, though the corporate giants will likely just pivot to complaining about 'safety' while their market share evaporates. It’s charming how quickly a billion-dollar lead can look like a legacy product when someone decides to stop hoarding the sauce.
Source: MiniMax
Comments
This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.