Qwen3-4B
Capacity estimate only; no automatic download or connection to your computer.
Model catalog & comparison
- Family
- Qwen3
- Weight size basis
- 4.02 B
- Documented/configured context limit
- 32,768
- License identifier (see source)
- apache-2.0
- Use
- Text chat
- Weight estimate GiB
- 2.33 · Q4_K_M
Compare under identical settings; size uses public tensor-element metadata and can differ from the named parameter count. Read licenses with base-model and usage terms.
Capacity scenario, not a benchmark: public weight tensor elements × bits ÷ 8 plus a 20% quantization/mixed-precision margin; F16 KV from layers, KV heads and context, plus 1.5 GiB runtime margin. Reserve at least 4 GiB or 25% RAM, and 2 GiB dedicated VRAM. One model, one text sequence; not a measured GGUF size or run guarantee.
Hybrid architectures, special packing or missing configs require manual review. Image/audio overhead is excluded. MoE uses total weights, not active parameters. 4/8-bit scenarios do not establish a published quant or backend compatibility.
Verified official quantized files
File sizes and hashes come from author repository listings; weights were not downloaded. File size is not total runtime memory. A verified matching-precision file takes precedence over the generic weight estimate, before cache and runtime margins.
- Q4_K_M · 2.33 GiB ↗
SHA256:
7485fe6f11af29433bc51cab58009521f205840f5b4ae3a32fa7f92e8534fdf5 - Q8_0 · 3.99 GiB ↗
SHA256:
8c2f07f26af9747e41988551106f149b03eb9b5cb6df636027b6bf6278473300
Models generate and reason. Agents use tools to work on tasks. A harness manages an agent’s environment. MCP connects tools and data. Skills are reusable task guidance.
Research notes
A growing local model snapshot, separated by chat, coding, retrieval, reranking and speech tasks. It spans generations, not the entire market or every latest release. Quality and speed are not uniformly benchmarked here.
Cost
Free software; check license and compute costs
Source code is public; model licensing, hardware, electricity and hosting are separate. Total cost is not verified.
Limits & boundaries
Hybrid architectures, special packing or missing configs require manual review. Image/audio overhead is excluded. MoE uses total weights, not active parameters. 4/8-bit scenarios do not establish a published quant or backend compatibility.
Original sources
- Qwen/Qwen3-4B ↗
huggingface.co · · Official documentation
- Qwen/Qwen3-4B-GGUF ↗
huggingface.co · · Official documentation