MiniMax has begun merging its M3 and H3 models into a single multimodal model, according to a report carried by MSN and sighted by this newsroom on 28 August 2026. The report presents the work as under way rather than finished, and provides no date by which the combined system is expected to ship.
The headline of that report is, at this point, close to the entire public record. It states that the laboratory is "beginning to merge M3 and H3 into one single multimodal model". No technical documentation, benchmark results, pricing or distribution details accompany the claim.
What the report says
Three claims can be extracted from the headline itself. First, that MiniMax maintains two distinct models designated M3 and H3. Second, that a process to combine them has started. Third, that the intended result is one model described as multimodal, a term the industry generally applies to systems that handle more than one kind of content rather than text alone.
The headline is phrased in MiniMax's own voice, which suggests it reproduces or summarises a company statement. The material available to this newsroom does not make the provenance explicit, so the claim is best treated as a company announcement relayed by a single outlet.
What the report does not say
The word "merge" is doing considerable work. It could describe training a unified architecture from shared components, wiring two existing systems together behind one interface, or consolidating two product lines under one name. Each reading implies something different about capability, serving cost and compute demand. The report does not distinguish between them.
The report also does not state which capabilities M3 and H3 each provide on their own, which modalities the combined model will cover, or how it will be licensed and distributed. No parameter counts, training data descriptions or evaluation results are given. There is no statement of motivation, whether competitive, commercial or research driven.
Notably, the report makes no reference to model distillation, training data provenance, or the wider disputes between laboratories over capability extraction. Nothing in the cited material connects this consolidation to those controversies, and readers should not infer such a connection.
Why the move is worth watching
Consolidation is a familiar point of interest when a laboratory operates several model lines, because combining systems can simplify deployment, share serving infrastructure and let strengths in one line carry over to another. Whether any of these considerations applies to MiniMax is not established by the report.
If the combined model does arrive as a single multimodal system, the questions that matter will be concrete: what it can do that M3 and H3 could not do separately, what it costs to run, and on what terms outside developers can use it. Until MiniMax publishes those details, any assessment of the merger's significance remains speculative.
Established versus claimed
- Established: an MSN report, sighted on 28 August 2026, states that MiniMax is beginning to merge its M3 and H3 models into one single multimodal model. Even this rests on one outlet's headline-level account.
- Claimed but unverified: that the merger is actively under way, and that the end product will be a single multimodal model. These points originate from what appears to be MiniMax's own framing and have not been independently confirmed.
- Unknown: the technical method of the merger, the modalities involved, the individual capabilities of M3 and H3, release timing, licensing, pricing, evaluation results, and any connection to the broader industry disputes this newsroom covers.