See what we shipped, how it measured, and how to run it.
Showing all 12 posts
The full ShapeLearn release of Qwen 3.8 27B: five GGUF models that all sit on the measured quality-speed frontier across six GPUs. GPU-5 is the default wherever it fits, at 99.63% of BF16, and GPU-4 stays very competitive when memory is tight. ShapeLearn-Lite held up better than its KLD ranking suggested, and we compare MTP against DFlash2 speculative decoding.
Read More →
An interactive explorer for the Krea-2-Turbo release: 24 curated prompts rendered by the GGUF and Humming builds, each family next to its own BF16 baseline. Compare any two variants side by side or with a slider, with synced pan and zoom.
Explore the comparison →
Our first image-generation release: quantized GGUF builds for ComfyUI and Humming builds for vLLM-Omni, generating a 1024×1024 image in roughly 8–9 seconds on a high-end GPU. Includes measured VRAM and speed for every model, setup guides for both runtimes, and 24 curated prompts compared across every variant of both families.
Read More →
A three-part series on quantized-model evaluation: does the model fit, how well does it perform, and how fast does it run on the target hardware? Why KLD measures displacement rather than direction, why BPW measures storage cost rather than realized speed, and why the quant that wins the proxies is often not the quant that wins in deployment.
Read More →
A Canada Day release: ByteShape-compressed GGUF models for Cohere North Mini Code 1.0, a Canadian coding model compressed by a Canadian team. See the best quality/speed trade-offs across RTX 4090, 5090, 4080, and 5060 Ti — pick Model 3 on 24GB+ GPUs, and Model 1 or 2 on 16GB GPUs.
Read More →
Compare standard and multi-token-prediction builds of Qwen 3.6 35B. Published tests show how model size, benchmark score, and token-generation speed vary across desktop and edge hardware.
Read More →
ByteShape's ShapeLearn-quantized release of Qwen 3.5 35B A3B. An MoE model where CPUs are surprisingly consistent but GPUs are much pickier. See the best quality/speed trade-offs across RTX 4090, 4080, 5090, RTX Pro 6000 Blackwell, Intel i7, Ryzen 9, Ultra 7, and Raspberry Pi.
Read More →
Step-by-step guide to running a fully local, fully free AI coding agent on your own hardware. Covers LM Studio, llama.cpp, and Ollama with ByteShape GGUF models, from installation to building Flappy Bird in one prompt.
Read More →
ByteShape's ShapeLearn-quantized release of Qwen 3.5 9B. GPUs agree on the best models, CPUs have strong opinions. See the best quality/speed trade-offs across RTX 5090, 4080, 3090, 5060 Ti, Intel i7, Ryzen 9, Ultra 7, and Raspberry Pi.
Read More →
Published comparisons of optimized Devstral and Qwen3-Coder builds on Raspberry Pi and RTX GPUs. Explore measured memory, speed, and benchmark-score tradeoffs, with download and setup links.
Read More →
A large language model running on Raspberry Pi 5, with published tests on Intel CPUs and NVIDIA GPUs too. Explore device-specific speed, memory, and benchmark-score tradeoffs, then find the build and setup instructions for your hardware.
Read More →
We're excited to announce ByteShape's first public release of ShapeLearn-quantized models. Learn how our datatype learning technology delivers better quality at lower sizes, with benchmarks across Qwen3 4B and Llama 3.1 8B models showing superior performance on GPUs, CPUs, and Raspberry Pi.
Read More →We haven’t published anything in this topic yet — it’s on the way. In the meantime, there is plenty else to read.