llama.cpp 0.4.0 adds four model architectures, an Apple RDMA transport, and video flags

llama.cpp ships nightly builds most days, and almost none of them are worth reading about. Version 0.4.0, published on September 4, 2026, is the other kind: it is a numbered release, it names four model architectures that would not load before it, and it changes two defaults that a running server already depends on.

Four architectures that would not load before

The headline entries are new families rather than new checkpoints. Qwen3.8-Flash-Next arrives under the internal name qwen4exp, and the release notes say the support is initial, with optimization still pending. That distinction matters: the weights load, but the speed they load at is not what the project intends to ship.

The other three are NVIDIA Nemotron-3-Puzzle-75B-A9B, DSpark support for Nemotron 3.5, and nanbeige4.2-3B. Alongside them, DeepSeek-V4 gained vision input handling and DeepSeek-V4-Flash-Vision-Exp was added, while Gemma-4 vision handling and Idefics3 preprocessing were fixed.

• Qwen3.8-Flash-Next, described as initial support with optimization pending.
• NVIDIA Nemotron-3-Puzzle-75B-A9B.
• DSpark support for Nemotron 3.5.
• nanbeige4.2-3B.

A new transport, and video reaches the command line

llama.cpp uses an RPC backend to spread one model across several machines, and 0.4.0 adds Apple RDMA as a transport for it, together with event and asynchronous backend APIs for RPC. On a desk with two Macs, that is a different wire under the same feature rather than a new feature.

Video input moved from library code to the command line. The release adds --video-* arguments, video and init options in the multimodal helpers, and a server that accepts data: URLs for video and audio input. The server also gained a per-slot context limit, so one long conversation no longer sets the ceiling for every other slot on the same process.

Two defaults changed, and saved state carries a new version

The quantizer now has a working-memory ceiling and streams rows in slabs, and model loading no longer spikes memory the way it did. Those three together are what let a large model be quantized on a machine that could not previously hold it, and they are the practical reason to update rather than the model list.

Three changes can break something that already works. Reasoning preservation is now on by default, the server rejects prefilled assistant messages that contain tool calls, and session and state versions were bumped for KV-cell token tracking. The notes list the last one under API changes without saying what happens to files written by 0.3.0, so anything that restores saved state should be tested before the update is treated as routine.

• Reasoning preservation is enabled by default, and the effective state is logged.
• The server refuses prefilled assistant messages carrying tool calls.
• ggml moved from 0.22.0 to 0.23.0, bringing sparse flash attention, asynchronous execution, and allocation-dependency tracking.