MiniMax Raises ~$2B While Its Frontier M3 Model Lands Inside FiftyOne

The raise

According to a dev.to post, Shanghai-based AI lab MiniMax has carried out a private placement plus convertible-bond raise reported at roughly HK$16 billion — on the order of $2 billion — its first major fundraising move since going public. The company listed on the Hong Kong Stock Exchange on 9 January at a valuation of about $6.5 billion, and its shares jumped 109% on the first day of trading, taking its market capitalisation above $13 billion. The post says the bulk of the new money is earmarked for model research and development and for infrastructure.

What shipped alongside: MiniMax-M3

The more consequential part for developers, the post argues, is the model that arrived with that momentum. MiniMax-M3 is a natively multimodal model — text, image and video are mixed in from the first training step rather than bolted on later — with roughly 428 billion total parameters, about 23 billion activated per token in a Mixture-of-Experts design, and a one-million-token context window.

It is built around what the company calls MiniMax Sparse Attention (MSA), a blockwise sparse attention mechanism described in its technical report. In that paper's controlled ablation on a 109-billion-parameter model, MSA matched standard Grouped Query Attention on downstream benchmarks while cutting per-token attention compute by 28.4x, with 14.2x prefill and 7.6x decode speedups at one-million-token context on H800 GPUs. The production M3 model card separately reports 9x prefill and 15x decode speedups over its predecessor M2 at the same context length. The post is careful to note that these two sets of figures come from different setups and are not interchangeable.

On public leaderboards, the post reports M3 scoring 80.5 on SWE-bench Verified, 59 on SWE-bench Pro, 78.1 on MMMU_Pro and 85.4 on Video-MME v2 — frontier-tier results on coding and multimodal reasoning benchmarks.

A plugin that turns M3 into a labelling tool

The practical hook of the post is minimax-m3-fiftyone, a plugin its author built that exposes M3 inside the FiftyOne computer-vision app. It loads as a FiftyOne Zoo model (foz.load_zoo_model("minimax/minimax-m3", task="detect")) and also appears as an App operator and a chat panel, requiring only a Hugging Face token — no fine-tuning and no model server to host.

It covers nine tasks end to end: detection, keypoint estimation, frame-level detection, temporal event search, key moments, two captioning modes, single- and multi-label classification, and visual question answering. Rather than returning raw text to parse by hand, it writes native FiftyOne label objects — Detections, Keypoints, TemporalDetections and Classifications — with coordinates normalised to the [0, 1] range.

There is no annotation-specific fine-tuning behind any of this; every grounded task is prompt engineering. The plugin asks M3 for JSON in an exact shape, and a deliberately forgiving parser cleans up the replies: it strips thinking blocks and code fences, scans left to right for the first balanced JSON span, tolerates alternative key names and wrapper objects, and degrades gracefully — a keypoint request that comes back as a bounding box uses the box's centre rather than dropping the item.

Video has no native upload path on the Hugging Face inference router, so the plugin decodes evenly spaced frames, timestamps them and sends them as an image sequence; temporal answers become TemporalDetection objects with start and end seconds. M3 has no audio modality on this path, and the post notes it returns an empty list rather than hallucinating a transcript when asked to transcribe speech — a limitation, but a reassuring failure mode.

A chat panel inside the sample modal streams multi-turn answers for each sample and can convert JSON-shaped replies into label fields with one click. Every label also carries its own provenance: the exact prompt, the raw model output, the thinking mode and the sampling parameters.

Why it matters

Two storylines intersect here. The funding one: a frontier lab that went public less than a year ago is still raising billions — roughly $2 billion in this round — and directing it at model R&D and infrastructure, evidence that the model race remains heavily capitalised. The distribution one: frontier models increasingly matter less as chat endpoints and more as components embedded in existing developer tools. M3 reaching FiftyOne through a community plugin, with no training and no infrastructure to stand up, shows how quickly a newly funded frontier model can land inside real workflows. The caveats are real, too — the grounding is prompt-driven rather than fine-tuned, outputs need a tolerant parser, and the headline speedup figures come from differing experimental setups. But the combination of deep funding and rapid tooling integration is precisely what keeps competitive pressure on every other lab in the field.