Choose for the task and available memory, rather than the largest parameter count. The first-run hardware check estimates memory suitability; it is not a promise of speed or successful use of every feature. The table below uses exact file sizes from this release’s model catalog. GB means decimal gigabytes; runtime memory and temporary space are extra.
| Model | Download total | Files | Publisher |
|---|---|---|---|
| Huihui Qwen3.8 Coder 27B · IQ3_S3 | 13.07 GB | 2 | Hugging Face ↗ |
| Qwen3.5 4B · Q4_K_M | 2.74 GB | 1 | Hugging Face ↗ |
| Qwen3.5 0.8B · Q4_K_M | 0.53 GB | 1 | Hugging Face ↗ |
| Qwen Image 2.1 Turbo · Q5_K_M | 11.21 GB | 3 | Hugging Face ↗ |
| Wan 2.1 T2V 1.3B · experimental | 6.75 GB | 3 | Hugging Face ↗ |
RAM and VRAM have different roles
RAM is system memory; VRAM is memory on the graphics card. Model weights, the context cache and inference buffers all consume memory. A download fitting on disk does not mean it fits in VRAM. Partial CPU offload can let a model run, but may reduce speed. More RAM helps when capacity is the constraint; it does not automatically make a GPU-bound workload faster.
A practical starting point
- A small 0.8B or 4B model is a reasonable first test on a low-memory computer. Expect more limited reasoning, coding and tool use.
- The standard Huihui Coder configuration targets around 32 GB system RAM and 16 GB NVIDIA VRAM. Leave at least 35 GB SSD space for the app, model packs and headroom; larger contexts need additional memory.
- Image generation uses the full three-file image pack. Switch between coding and image generation to avoid keeping both large workloads in VRAM. GPU support and drivers still matter.
What does quantization change?
GGUF quantization reduces weight size. Lower precision may allow a larger model to fit, with a possible quality cost. Q4 and IQ3 names are not interchangeable quality guarantees across models. Compare outputs on your own tasks and keep the publisher’s model card and license.
Context and tokens per second
A longer context consumes more memory and can slow processing. Saving project notes and summarizing history helps manage the window, but cannot guarantee that every detail is preserved. A published long coding test on Windows 11 with an RTX 4080 measured 100.85 tokens/s on its specific setup. It is not a promise for every PC or a full 65K conversation.
Images, photos and experimental video
Image generation and photo analysis require different compatible model components. Do not infer visual capability from the word “Coder”. Wan Video Studio is experimental: a Windows RTX 4080 test generated a 512×288, 33-frame, 20-step clip in 52.68 seconds including restoration of the coding model. Other hardware and operating systems remain unverified.
Sources and verification
File sizes and download hashes come from the published catalog. Read verification notes and limits before planning production work. Models are downloaded from their publishers and retain separate licenses.