Make videos on the computer you already own. FreeVideo runs MiniMax H3 in as little as 8 GB of VRAM and 16 GB of RAM, and adapts its acceleration path to your hardware.
Commits on main
Head groups share one LoRA down projection per layer call instead of repeating it for every group
08c9257 Bowen Xue committed 25 days ago
Keep the full recovery streak of a GPU-error level when unexplained interruptions raise it further
370c150 Bowen Xue committed 25 days ago
One completed generation restores the level raised after an interruption of unknown cause
e8c14e5 Bowen Xue committed 25 days ago
Linear-branch fused kernels compile once for every canvas, clip length and head split
f0c0198 Bowen Xue committed 26 days ago
Keep an idle engine when its own pinned weights already cover the next request's host cache
afba5fc Bowen Xue committed 25 days ago
Keep the model warm after a successful video until something needs the memory, with no idle limit
3ec8d4c Bowen Xue committed 26 days ago
Keep-warm release: hold the start lock and free the card in order
68fe2cc Bowen Xue committed 27 days ago
Keep the model warm only when the card's use can be read; drop the unused release parameter
ba3de55 Bowen Xue committed 27 days ago
Audio cards: Use as is goes back to Reference under a preview first or a continued video
c187dd3 Bowen Xue committed 23 days ago
Audio references: a clip can be used as is for the video's sound
d73e253 Bowen Xue committed 23 days ago
Keep the pass mode in step when an example sets two-pass and preview
e36822e Bowen Xue committed 23 days ago
Add the feature tour, first-run guide and tutorial try-it entry
6be4b9d Bowen Xue committed 26 days ago
Write the tutorial book chapters with pictures and a Try it button
44dc88c Bowen Xue committed 23 days ago
Merge pull request #417 from FlashML-org/release/0.4.18-public
0afd01e Bowen Xue committed 1 month ago
Give the full-disk failure reason an English line on Autopilot cards
13c7a9f Bowen Xue committed 1 month ago
Launcher: a missing prompt model list no longer fails the whole page
0848a94 Bowen Xue committed 1 month ago
Let a Windows request spend its default RAM growth reserve before the planner refuses it
5001732 Bowen Xue committed 1 month ago
Placing a picture from the image history keeps the generation tool, so the history stays visible
5a8ecfc Bowen Xue committed 1 month ago
Check MLX image attention against Torch once and report why it fell back
ff735cd Bowen Xue committed 1 month ago