Skip to content

Paddock 0.1.15

Choose a tag to compare

View all tags
@panterlo panterlo released this 09 Oct 06:02
v0.1.15
6e2d6f1

A fixes and polish release: the RAM context cache no longer switches itself
off when it fills up, Claude Code sessions on Kolibri 1 run longer and
faster, tool definitions reach every model exactly as its chat template
expects, and on the Mac EmbeddingGemma 2 now takes pictures and audio.
Windows x64, Linux x64 and the NVIDIA DGX Spark. NVIDIA GPUs, driver 580 or
newer. The macOS pre-release for Apple Silicon is built from the same commit.

New

  • TIC Forestry v1 (The Intelligence Company). Land cover and canopy
    height at 1 m from Swedish aerial imagery, near-infrared included, in one
    pass. Over the API, /v1/segmentations. Like SAM 3 it is a gated
    download: accept the DINOv3 licence on its Hugging Face page and add a
    token in Manager > Settings. On NVIDIA GPUs.

  • Gated downloads are marked in the model picker with a lock and
    "licence on Hugging Face", so you know before you start that the download
    needs your own token.

Improved

  • Kolibri 1 in long agent sessions. Reading long prompts and decoding
    deep into the context are faster with the 8-bit context cache, the
    default (Nemotron's long-context decode gains too), and so is the NVFP4
    version's prompt reading on Blackwell. After a reply that calls a tool,
    the next request resumes from the cache instead of reading the reply's
    reasoning again.

  • Tool definitions reach the model as its chat template writes them.
    Tool schemas and earlier tool calls kept their keys in alphabetical order
    and had <, >, & and apostrophes escaped; they now arrive exactly as
    the reference renderer writes them, on every model family.

  • System messages in the middle of a conversation stay where they are
    on models whose templates allow it - Kolibri 1, Laguna, Gemma 4, Granite
    and Muse Glimmer - instead of being folded into the next user turn.
    Claude Code ends every request with one.

  • Faster pictures on Gemma 4 and EmbeddingGemma 2, which share a picture
    tower, and faster EmbeddingGemma 2 text embeddings on NVIDIA GPUs. The
    results are the same to the bit.

  • Kumo Tabular asks Size and Task on the start form instead of offering
    six quality cards, and the Tables page offers to start the other task -
    predicting numbers instead of classes, or the reverse.

Fixed

  • The RAM context cache no longer goes offline when it fills up. A full
    cache could have free space and still no gap big enough for the next
    block; those stores counted as failures, and enough of them switched the
    cache off until the model restarted ("Cache offline" in the Studio). A
    full cache now makes room before it stores, and a cache that is simply
    full is never switched off.

  • Claude Code's thinking display setting is accepted by the Messages
    API instead of refusing the request.

macOS (pre-release)

  • EmbeddingGemma 2 takes pictures and audio on Metal, from the GGUF with its
    picture and audio downloads or from an MLX 8-bit package that holds both;
    video comes in as sampled frames, without the soundtrack.
  • The native app has an Embeddings workspace.
  • MLX text embeddings no longer depend on what else is in the batch.

Known

  • SAM 3 and TIC Forestry run on NVIDIA GPUs only, and their weights
    download only with a Hugging Face account that has accepted the model's
    licence. TIC Forestry is built for Swedish imagery at 0.5 m; it is not a
    general-purpose aerial model.

  • Kolibri 1 has no same-weights llama.cpp reference yet: no llama.cpp
    release reads it, so its correctness is checked against Aleph Alpha's own
    serving code.

  • EmbeddingGemma 2 takes audio clips of up to 30 seconds; on NVIDIA GPUs
    video is not served yet.

  • Laya reads text only, and its confidence on choices with more than ten
    options is not calibrated: the English checkpoint ships an out-of-range
    temperature for that case, which Paddock clamps, as Laya's own server does.

  • Kumo Tabular predicts up to ten classes.

  • On Qwen 3.8 Flash-Next's NVFP4 versions a greedy speculative reply can
    part from plain decoding at a near tie     - the same model with slightly
    different numerics, not a wrong answer.

  • On-demand loading covers Whisper only.

  • The fp8 KV cache's paged attention on RTX 50-series cards shows a small
    numeric deviation in one split configuration (#5).

Paddock 0.1.15, built from 6e2d6f18 on the maintainers' machines (Windows natively, Linux x64 in a Docker builder, the DGX Spark natively on a Spark), each archive booted on a GPU and asked a question before it was uploaded.

Requirements

  • An NVIDIA card the kernel pack covers - x64: RTX 30/40/50 series, the RTX A / PRO line, B200; aarch64: the DGX Spark (GB10). Other cards get an honest refusal, never a slow fallback. This build: linked in (sm_86 sm_89 sm_100 sm_100a sm_120 sm_120a), nvcc 13.3.73.
  • A driver from the r580 branch or newer. No CUDA toolkit: the kernels and the runtime are inside paddock-runner.
  • Windows 11 x64; Linux x64 with glibc 2.35 or newer (Ubuntu 22.04+, Debian 12+); or a DGX Spark on DGX OS 7 (Ubuntu 24.04, glibc 2.39).

Run

  • Unzip or untar anywhere. paddock starts the manager and the Studio on http://localhost:11500. paddock-runner --model <file.gguf> serves a bare OpenAI/Anthropic-compatible endpoint. README.txt inside says the rest.

sha256

3c82ba04d2652f26c6441c62bbc909072d68d2a1b8bdb047617f730019a89fd9  paddock-0.1.15-x86_64-windows.zip
5e522829528ff4eb94d3d42ff38792992f3615326f9ac14b473660bb4ae0ada8  paddock-0.1.15-x86_64-linux.tar.gz
c20fa46f4bfbafaafac40f5de8d227ec76bc2faa2d78580d07633078c1c3c48a  paddock-0.1.15-aarch64-linux.tar.gz
Assets 16
Loading