MiniMax H3 team held an AMA on Reddit, addressing developers' hot questions about open-sourcing the 2K model, sparse attention, and image generation. The team confirmed plans to release the 2K model and provide a sparse attention reference implementation, while considering Apache-2.0 licensing.
SynthePulse Insight · AI deep readingMembers
Version 1 · 1 source
In a Reddit AMA, the MiniMax H3 team gave their first concrete responses about what's next after open-weight release: a 2K regeneration model is coming soon, a sparse attention reference implementation is in the works, low-step versions are under consideration, and a dedicated image generation model has been derived from the H3 lineage. But Context-IR remains API-only, and Apache-2.0 is still just 'under consideration.'
One of the community's biggest questions is whether H3's 2K capability can be reproduced locally. MiniMax's developer relations lead, Ryanlee, answered clearly: 'Yes, and it won't be long.' Researcher Kiro further explained that H3-Regenerate-2K is essentially a second-stage conditional generation process, using a latent-space DiT regeneration checkpoint specifically designed for higher target resolutions. The base model's generated content is fed as additional context, with some reference inputs provided at higher resolution. This is not simply rerunning the Base Checkpoint, nor is it a traditional upscaler.
This statement means that after the open-weight release, the community's inability to fully reproduce the online 2K capability will change, but no specific timeline was given.
This article is based on AI Front's coverage of the Reddit AMA; all information comes from MiniMax team's public answers in the AMA, which are source_claim level. Some details (such as specific timelines, model performance) lack first-party verification and should be treated with caution.
Primary report
Primary source