Qwen3-Coder 30B-A3B: how much RAM do you need?
Qwen3-Coder 30B-A3B: butuh RAM berapa?
About 17 GB for the weights at 4-bit, and about 19 GB in total with an 8K context (21 GB at 32K). Numbers are estimates.
Sekitar 17 GB untuk bobot pada 4-bit, dan sekitar 19 GB total dengan konteks 8K (21 GB pada 32K). Angkanya perkiraan.
Specs
Spesifikasi
| Parameters | Parameter | 30.5B |
|---|---|---|
| Layers / KV heads | Layer / kepala KV | 48 / 4 |
| Native context | Konteks asli | 256K tokens |
| License | Apache 2.0 |
Memory by context length
Memori menurut panjang konteks
| Context | Konteks | KV cache | Q4 | Q8 | FP16 |
|---|---|---|---|---|---|
| 4K | 0.4 GB | 18 GB | 32 GB | 58 GB | |
| 8K | 0.8 GB | 19 GB | 32 GB | 59 GB | |
| 16K | 1.5 GB | 20 GB | 33 GB | 59 GB | |
| 32K | 3.0 GB | 21 GB | 34 GB | 61 GB | |
| 64K | 6.0 GB | 24 GB | 37 GB | 64 GB | |
| 128K | 12 GB | 30 GB | 43 GB | 70 GB | |
| 256K | 24 GB | 42 GB | 55 GB | 82 GB |
Totals include the weights, the KV cache and about 1 GB of runtime overhead. The KV cache column shows the memory the context adds. Real use varies with the app and settings.
Total mencakup bobot, KV cache, dan sekitar 1 GB beban runtime. Kolom KV cache menunjukkan memori tambahan dari konteks. Pemakaian nyata bergantung pada aplikasi dan pengaturan.
Good to know
Perlu diketahui
- Text only: it cannot read images. Use a vision model for that.
- Mixture of experts: all weights must fit in memory, but it runs faster than its size suggests.
- Hanya teks: tidak bisa membaca gambar. Pakai model vision untuk itu.
- Mixture of experts: semua bobot harus muat di memori, tapi jalannya lebih cepat daripada ukurannya.
Source: Hugging Face model card · checked 2025-12-03 (model files), page built from the model's configdicek 2025-12-03 (berkas model), halaman dibuat dari konfigurasi model