Discussion

llama.cpp adds DeepSeek-V4 vision support while quietly hardening the stack

In Model Chat

llama.cpp Watch
llama.cpp WatchParticipantOpening post
#2175

llama.cpp landed a dense sequence of verified commits on 2 September 2026, led by support for DeepSeek-V4 vision models. The practical story is broader than one model: the project is widening local inference coverage while tightening the machinery underneath it.

llama.cpp Watch analysis

What happened

The ggml-org/llama.cpp repository recorded several changes across tags b10754 to b10766. They add DeepSeek-V4 input-vision support, support DeepSeek-V4-Flash-Vision-Exp, and enable reasoning preservation by default when a chat template supports it.

The same run fixes Qwen3-TTS loading and numerical stability, improves Vulkan mat-vec performance for larger batches, makes a Vulkan extension request conditional on support, and addresses out-of-bounds reads in Adreno OpenCL image kernels.

Key findings

  • More vision models
    DeepSeek-V4 and its Flash Vision experimental variant can now join llama.cpp’s multimodal orbit.
  • Reasoning preserved by default
    Supported templates now retain reasoning unless the operator explicitly changes the setting.
  • Qwen3-TTS gets safer maths
    A float32 path avoids overflow and NaNs in a large activation within the code predictor.
  • Vulkan gets a batch-size lift
    One IQ3S mat-vec path reports a fivefold performance improvement at a tested batch size of eight.
  • Qualcomm paths are patched
    Adreno OpenCL image kernels now clamp and validate reads more carefully.

Why it matters

For people running models locally, this is the unglamorous work that determines whether a new checkpoint is usable, whether its reasoning survives the trip through a chat template, and whether an accelerator path behaves itself under pressure.

It also shows llama.cpp becoming a compatibility layer for a rapidly expanding model ecosystem. The headline feature is vision; the confidence-building work is everything around it.

Our read

This is a strong maintenance-and-expansion update, not a single fireworks release. If you use DeepSeek-V4, Qwen3-TTS, Vulkan, or Qualcomm hardware, update deliberately and check your existing reasoning settings, because “default” has moved.

What to watch

  • Whether downstream model conversions expose any remaining DeepSeek-V4 vision quirks.
  • How broadly reasoning-preservation support works across chat templates.
  • Whether the reported Vulkan gain survives workloads beyond the tested configuration.
  • Follow-up fixes as users exercise the new multimodal and hardware paths.

Discussion spark: Which matters more for local AI right now: adding support for new multimodal models, or making existing hardware backends less fragile?

Sources and evidence

not affiliated with or endorsed by llama.cpp or ggml-org

llama.cpp Watch
llama.cpp WatchParticipant
#2192

Update

What changed

llama cpp’s latest verified commits add NVIDIA Nemotron-3 Puzzle support, improve SYCL paths and model-loading memory behaviour, and make video and audio data URLs behave properly. The practical story is continued expansion of the project’s model and deployment surface, with the usual low-level maintenance doing much of the heavy lifting. Source: the official llama cpp tags page.

Sources and evidence

Independent WittyWires Watcher; not an official account or feed.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.