Disclosure: As an Amazon Associate I earn from qualifying purchases. This site contains affiliate links.
llama.cpp 0.3.0 Is a Tag, Not a GX10 Bench
ggml-org tagged llama.cpp v0.3.0 on Aug 25, 2026. Nightly build on that release is b10621. ggml is bumped to v0.22.0.
Key takeaways
- ggml-org tagged llama.cpp v0.3.0 on Aug 25, 2026. Nightly build on that release is b10621. ggml is bumped to v0.22.0.
- The tag adds dots3-note (DSA-ISWA KV, vision/audio in mtmd), MTP for GLM-4.5-Air, DeepSeek 4 -sm tensor, WebP via ffmpeg, and a fix for videos whose moov atom sits at the end of the file.
- dots-studio's card: Apache 2.0, 280B total / 16B active, text+image+video+audio. Their vLLM example is eight H100s at TP=8. I have not timed any of this on the GX10.
Local LLMs on NVIDIA Spark / ASUS GX10
ggml-org tagged llama.cpp v0.3.0 this morning. The instrument is a GitHub release, published Aug 25 at 10:22 UTC. Nightly on that page: b10621. ggml is v0.22.0.
Headline items from the notes, not from my bench: dots3-note with a new DSA-ISWA KV cache, MTP for GLM-4.5-Air, DeepSeek 4 tensor-split via -sm tensor plus a multi-sequence rollback fix, bailingmoe3 DSpark. mtmd gets dots3-note vision and audio, WebP through ffmpeg, a Pillow-accurate resize, and a fix for videos whose moov atom sits at the end of the file. Server gets LLAMA_SERVER_SLOTS_N_DIFF. Web UI gets tabbed chat.
The dots3 card is dots-studio/dots3-note-prev. Apache 2.0. 280B total, 16B active, 512K context, text+image+video+audio. Their own vLLM example is the FP8 checkpoint on eight H100s at --tensor-parallel-size 8. That is not a GX10 sit. I am not going to pretend the tag made it one.
The useful local line is smaller. The moov-at-EOF video fix is why a 27B vision+video sit was parked. WebP decode is real. DeepSeek -sm tensor is a multi-GPU flag. None of those are numbers from this box.
I run Grok when it earns it, and a GX10 when I want the weights in the room. This tag is the software side of that split. If I upgrade and time something, that note comes next. Not this changelog.
Frequently asked questions
Yes. ggml-org tagged v0.3.0 on Aug 25, 2026 at 10:22 UTC. The release notes list dots3-note, GLM-4.5-Air MTP, DeepSeek 4 tensor split, and mtmd video/WebP fixes. The nightly build on that page is b10621.
Not from anything I have sat. The Hugging Face card for dots-studio/dots3-note-prev is Apache 2.0, 280B / 16B active, 512K context. The official vLLM recipe deploys the FP8 checkpoint on eight NVIDIA H100 GPUs with tensor-parallel-size 8. That is an 8-GPU node, not this box.
No. It unblocks loading videos that put the moov atom at the end of the file, which is why that sit was parked. I have not upgraded this machine or timed the 27B. A changelog is not a tok/s.
Affiliate Disclosure: As an Amazon Associate I earn from qualifying purchases. This site contains affiliate links.