ByteShape's low-VRAM quantizations of Alibaba's multimodal Qwen3.8-27B. The recommended build runs in about 13GB of VRAM on a 16GB consumer GPU, still reads images, and reportedly keeps ~99.6% of the full model's benchmark score. Apache 2.0.
Last verified · 2026-09-18 · by Moe Ameen
ShapeLearn Qwen 3.8 27B is not a new model — it is a set of quantized GGUF builds of Alibaba's Qwen3.8-27B, produced by ByteShape, a small outfit whose tagline is "AI models, shaped for real hardware." The base model is the dense, single-GPU, multimodal member of Alibaba's Qwen3.8 line: a roughly 27-to-28-billion-parameter LLM with a vision encoder that reads images and video, a long native context, and configurable reasoning, released under Apache 2.0 in August 2026. What ByteShape does is make it dramatically cheaper to run.
The method is called ShapeLearn, and it is the reason this release matters more than a routine repack. Instead of applying one uniform bit-width across the whole network, ShapeLearn learns the optimal numeric datatype for each tensor — spending precision where the model is sensitive and stripping it where it is not. The practical result is that the builds hold quality at bit-widths where a blunt quant would degrade. ByteShape published a "ShapeLearn-Lite" preview in mid-August 2026 and the full ShapeLearn GGUF set on Hugging Face on September 14, 2026. The release ships five GPU-optimized variants, labeled GPU-1 through GPU-5, from roughly 2.56 to 3.84 bits per weight — about 8.8GB to 13.1GB of weights.
The headline number is the recommended GPU-5 build: about 13GB of VRAM, which fits a 16GB consumer GPU and leaves headroom on a 24GB card, while ByteShape reports it retaining around 99.63% of the original BF16 model's aggregate score across math, coding, general-knowledge, instruction-following, and agentic benchmarks. That is a different hardware bracket than the roughly 28GB the official FP8 build required. Two more things carry over: the multimodal capability survives compression through llama.cpp's MTP speculative-decoding path (the vision projector downloads automatically), while the alternative DFlash2 path is text-only but faster, and speculative decoding adds meaningful throughput on top, on the order of 1.3x to 2x depending on method. The builds inherit Qwen3.8-27B's Apache 2.0 license. Confirm the per-variant sizes, benchmark figures, and llama.cpp build requirements on ByteShape's model card before you commit to a specific number — this is a fast-moving community release and figures were still settling at publication.
The thing worth naming about a ShapeLearn build is what it changes economically: once the model is loaded on a card you already own, drafting is effectively free and offline. You can run it on a plane, off a home desktop at 2am, or on a laptop GPU without touching an API meter — and it still sees, so it will read your rough footage and screenshots as easily as it reads a prompt. That unlocks a working style hosted models discourage: draft aggressively. Generate forty hook variants for a launch, have it watch a raw cut and pull the three strongest moments, rewrite a script five different ways — none of it costs a token, so you stop rationing. The catch is that all of that output is still just text and analysis sitting in a terminal. The model renders no video, designs no image, and posts to nothing.
That is exactly the seam [Kompozy](/) fills. Point Kompozy at your local llama.cpp endpoint with bring-your-own-key on the Founding tier, and the volume you just drafted for free becomes finished, on-brand media: the winning hooks become [Persona Shorts](/glossary/persona-shorts) talking-head videos and [Marketing Shorts](/glossary/output-buckets), the moments it flagged in your footage become captioned [Clipped Shorts](/glossary/content-repurposing) reframed to 9:16, 1:1, and 16:9, the best angles become brand-exact [Carousel Posts](/glossary/output-buckets) and [Quote Graphics](/glossary/output-buckets) rendered through [HyperFrames](/glossary/hyperframes), and the long-form skeletons become a [Blog Article](/glossary/output-buckets) and an email newsletter — every one held to your voice by the [Persona Brief](/glossary/persona-brief) and banned-word filters, then scheduled and published across the eight social platforms plus blog and email on [Autopilot](/glossary/autopilot). The local model gives you cheap, private raw material at scale; Kompozy turns it into a published week. (Kompozy's own copy generation runs on managed Claude and OpenAI, so a local ShapeLearn model is a private drafting-and-vision brain feeding the pipeline, not a swap for it.)
It is a set of quantized GGUF builds of Alibaba's Qwen3.8-27B, produced by ByteShape using its ShapeLearn method, which learns the numeric datatype for each tensor rather than applying one uniform bit-width. The full set landed on Hugging Face in September 2026. The recommended build fits in about 13GB of VRAM, still reads images, and reportedly keeps about 99.6% of the original model's benchmark score. It is Apache 2.0.
The five variants span roughly 8.8GB to 13.1GB of weights, so the recommended ~13GB build fits a 16GB consumer GPU and runs comfortably on a 24GB card like a 4090 or 3090. That is well below the ~28GB the official FP8 build needed. Budget extra VRAM for context and, on the DFlash2 speculative path, its separate draft model, and pick the largest variant your card holds.
Yes, through llama.cpp's MTP speculative-decoding path, which downloads the vision projector automatically. The DFlash2 path is faster but text-only. As with the base model, vision is an input capability: it reads images and video you give it but generates no media of its own.
No. It drafts and reasons over text and reads images and video, but it produces no finished video, images, or designs, enforces no brand voice, and publishes to no platform. To turn its drafts and analysis into on-brand posts across platforms, pair it with a content engine like Kompozy — which can call your own local Qwen endpoint on the Founding tier — that renders the media and handles scheduling and publishing.