Qwen3-0.6B Q4_K_M needs 1.81 GiB at 8,192 tokens
Qwen3-0.6B-Q4_K_M.gguf (unsloth/Qwen3-0.6B-GGUF) is 396,705,472 bytes (0.37 GiB, 0.40 GB). The auto-selected artifact is GGUF Q4_K_M (Unsloth). At 8,192 tokens, 1 sequence, FP16 KV, llama.cpp, the calculator estimates 1.37 GiB–1.81 GiB (1,470,447,296–1,946,801,276 bytes). GGUF Q8_0 is 639,446,688 bytes. At 8,192 tokens on one device, RTX 4060: 1.37 GiB–1.81 GiB, Runs comfortably, net budget 7.00 GiB. RTX 4070 Ti SUPER: 1.37 GiB–1.81 GiB, Runs comfortably, net budget 15.00 GiB.
- Qwen3-4B Q4_K_M VRAM
- Qwen3-1.7B Q4_K_M VRAM
- Qwen3-8B Q4_K_M VRAM
- RTX 4060: Runs comfortably
- RTX 4070 Ti SUPER: Runs comfortably
- Qwen3-8B Q4_K_M on RTX 4060
Specific checkpoints with sourced architecture and weight files. GGUF Q4_K_M (Unsloth): 1.37 GiB–1.81 GiB at 8,192 tokens, 1 sequence, FP16 KV, llama.cpp.
Publisher and provenance
- PublisherQwen
- Hugging Face
Qwen/Qwen3-0.6B - Revision
c1899de289a04d12100db370d81485cdf75e47ca - Total parameters751,632,384
- Layers28
- Query heads16
- KV heads8 (gqa)
- Head dimension128
- Recorded context capability40,960
- Completenesspartial · full
Memory layers in this estimate
Artifact bytes, raw KV, engine extras, and total demand are separate. This is not a runtime trace.
- Weight artifact
- 0.37 GiB (396,705,472 bytes)
- Raw KV cache
- 0.88 GiB (939,524,096 bytes)
- Engine extras / workspace
- 0.13 GiB–0.57 GiB
- Total estimated demand
- 1.37 GiB–1.81 GiB
Published artifacts
| Quantization | Selected files | Bytes | Engine | Evidence |
|---|---|---|---|---|
| GGUF Q8_0 (Qwen) gguf-q8_0 | Qwen3-0.6B-Q8_0.gguf | 639,446,688 (0.60 GiB / 0.64 GB) | llama.cpp, ollama, unspecified | file-metadata |
| GGUF Q4_K_M (Unsloth) gguf-q4_k_m | Qwen3-0.6B-Q4_K_M.gguf | 396,705,472 (0.37 GiB / 0.40 GB) | llama.cpp, ollama, unspecified | file-metadata |
| GGUF Q5_K_M (Unsloth) gguf-q5_k_m | Qwen3-0.6B-Q5_K_M.gguf | 444,415,680 (0.41 GiB / 0.44 GB) | llama.cpp, ollama, unspecified | file-metadata |
| GGUF Q6_K (Unsloth) gguf-q6_k | Qwen3-0.6B-Q6_K.gguf | 495,107,776 (0.46 GiB / 0.50 GB) | llama.cpp, ollama, unspecified | file-metadata |
| INT4 theoretical | No published file | 375,816,192 (theoretical) | — | theoretical |
Memory by quantization and context
Concurrent sequences = 1. KV defaults to FP16. The range includes documented runtime extras for the listed engine. These figures use the same core as the interactive calculator.
| Quantization | 2,048 tokens | 8,192 tokens | 32,768 tokens |
|---|---|---|---|
| GGUF Q8_0 (Qwen) | 0.94 GiB–1.33 GiB | 1.60 GiB–2.04 GiB | 4.22 GiB–4.85 GiB |
| GGUF Q4_K_M (Unsloth) | 0.71 GiB–1.10 GiB | 1.37 GiB–1.81 GiB | 3.99 GiB–4.62 GiB |
| GGUF Q5_K_M (Unsloth) | 0.76 GiB–1.15 GiB | 1.41 GiB–1.86 GiB | 4.04 GiB–4.67 GiB |
| GGUF Q6_K (Unsloth) | 0.80 GiB–1.19 GiB | 1.46 GiB–1.91 GiB | 4.09 GiB–4.72 GiB |
| INT4 theoretical Theoretical rows are not downloadable artifacts. | 0.69 GiB–1.28 GiB | 1.35 GiB–2.10 GiB | 3.98 GiB–5.33 GiB |
Raw KV contribution
Standard GQA/MHA raw KV is 2 × layers × kvHeads × headDim × tokens × sequences × bytes per element. This page uses KV heads (8), not query heads (16). Using query heads would overstate this cache by 2×.
- 8,192 tokens, 1 sequence, FP16: 939,524,096 bytes (0.8750 GiB)
- 32,768 tokens: 3,758,096,384 bytes (3.5000 GiB)
Related pages
Sources
- Qwen3-0.6B model card
- config.json — revision c1899de289a04d12100db370d81485cdf75e47ca
- Artifact repo Qwen/Qwen3-0.6B-GGUF
- Artifact repo unsloth/Qwen3-0.6B-GGUF
Assumptions that affect this result
- Requested context is within the recorded native window of 40960 tokens.
- Full-resident weight estimates use total parameters, including inactive MoE experts.
- Weight bytes come from the selected weight files only, not from every file in the repository.
- Download file size is not peak GPU memory. Runtime layout, allocator padding, and KV are extra.
- Counted files: Qwen3-0.6B-Q4_K_M.gguf.
- Context budget is the number of cached tokens per sequence (prompt plus reserved generation).
- Concurrent sequences share one model replica; weights are not multiplied by sequence count.
- KV dtype is independent of weight quantization unless you change it.
Calculator
Loading catalog snapshot…
Search or pick a catalog checkpoint to see memory demand.