So far, this release is an experiment on MiniMax H3 base model that combines learned patterns from two other video diffusion models — LTX 2.3 and Wan 2.2, and in test3 and later, one image diffusion model — Krea 2 — into H3's transformer architecture.
A unified transformer with modality-specific input projectors (video, audio, condition) feeding into a 52-block transformer stack (2-block token_refiner for text conditioning plus 50 main blocks), with modality-specific output heads. The transformer stack uses identical block architecture throughout, which is what makes the attention and MLP layers architecturally-appropriate targets for character transfer between models — modality routing lives at the pipeline edges, not within the transformer
The LTX pass was applied first to H3's front blocks, followed by a Wan pass targeting H3's middle-to-back blocks. Each pass modified attention weights directly; the second pass also modified the feed-forward layers in a broader block range. Together, these two video-donor passes demonstrated that completely different video model architectures could each contribute their own recognizable character into H3.
A third pass then brought in Krea 2 — an image diffusion model rather than a video model. This was a more speculative direction, since image models learn spatial relationships within a single frame while video models learn spatiotemporal relationships across many frames. Krea's contribution was scoped narrowly to the parts of its attention structure that carry that spatial learning: the query projection (which encodes what visual features to attend to) and, in a smaller number of positions, the paired key and value projections (which encode how those features relate). The feed-forward layers were partially transferred as well, but only the input side of Krea's gated MLP — the output projection was skipped because Krea's combining logic was trained for an image-model output pathway that doesn't match H3's video processing. This narrower scope let Krea contribute finer visual character and detail work without disrupting H3's temporal coherence.
MiniMax's own architecture documentation notes that H3's attention and feed-forward layers do not contain modality-specific structure — the parts of the model that handle "this is audio vs video vs text" live elsewhere in the model. That means the attention and feed-forward weights are the architecturally-appropriate location to transfer character from other trained models without disrupting how H3 handles each modality. This applies equally to same-domain video donors and to cross-domain image donors, though cross-domain transfer requires more careful scoping since not every part of an image model's learned patterns has a coherent home in a video model.
The modifications use orthogonal projection math — new weight patterns are added into H3 primarily in directions the base model wasn't already using. This lets the transferred character supplement H3 rather than override it. For the cross-domain Krea pass, this was especially important: Krea's patterns and H3's patterns turned out to be nearly perpendicular in weight space, giving orthogonal projection maximum room to work while preserving H3's own directions almost entirely.
Combined with careful control over which blocks are affected — and, for Krea specifically, careful control over which attention head positions receive the injection to avoid disturbing the head slots most involved in audio processing — this preserves H3's core capability to produce coherent video with audio while shifting its aesthetic and motion character toward the donor models. The model's actual block weights, safety features, and guardrails remain unmodified in the architectural sense — every modification is a controlled additive perturbation of existing weights, not a replacement of the model's own structure.
In the final beta1 configuration, the H3 model received surgical grafts from other donors targeted at the token_refiner blocks 0 and 1 — the two-block text refinement stack that processes Gemma text embeddings before they condition main-block generation. Because token refinement shapes what the main blocks receive as text conditioning, changes here influence the entire prompting direction and output structure.
The H3 model's fused Q/K/V projections were unfused into separate q_proj, k_proj, v_proj tensors — necessary because the fused format doesn't allow per-band surgery, but does require refusing them back into the fused qkv_proj format for ComfyUI inference.
attn_q was given to Krea to shift the token refinement's "what to attend to" direction toward Krea's learned patterns, altering styling at the token level.
attn_k was isolated. This is worth naming as an empirical finding for anyone training or merging into H3: attn_k appears to carry critical calibration for the model's amplification behavior. Even slight perturbation degrades audio quality along with other finely-trained details from the model's reinforcement learning. Since H3 doesn't have explicit control gates (unlike LTX which uses to_gate_logits), it appears to rely on precisely-tuned attn_k to modulate attention strength. Standard training pipelines that hit K without knowing this may be silently degrading the model. Isolating K requires unfusing the qkv triplet, targeting only Q and V, and refusing — which the beta1 toolchain supports.
attn_v was given to LTX-2.3 (Eros lineage) via linear-mag mode graft. V carries a substantial portion of the output character, so this was where LTX Eros content flows into H3's text refinement.
The triplets were refused back to fused qkv_proj format, and then Wan 2.2's main-block MLP fc1 was grafted into blocks 0 and 1 as a supplementary contribution — Wan 2.2's dimensions align nearly perfectly with H3's token_refiner MLP inner (96% coverage), making it structurally well-suited.
Together these grafts inject a substantial but subtle influence: they bring specific motion details across, adjust t2v styling, and do this without removing prompting or behavior capability from the base model in the way many training and merging approaches do. Some audio quality degradation is present but limited. Given H3's unified audio-video attention architecture, meaningful modification of the video path is difficult to separate from audio impact — the unification is genuinely difficult to work around at the merge/graft level.
The standard H3 community license applies. Because this release now carries transferred character from LTX 2.3, Wan 2.2, and Krea 2, the community licenses for those source models apply as well to the portions of character that came from each.