Qwen

Alibaba's Qwen (通义千问) is the largest open-weight family in our catalog: 22 models spanning four generations — Qwen2.5, Qwen3 (including the 2507 refresh), Qwen3.5 and Qwen3.6 — plus dedicated Coder, Thinking and Next branches. Every one of them ships under the Apache-2.0 license, context windows run from 128K up to 256K, and the range covers everything from a 2B pocket model to a 480B MoE coding flagship.

All VRAM figures below use our standard estimate: Q4 weights + KV cache at 8K context + 1 GB overhead. For MoE models, the KV cache tracks the active parameters, not the total — which is exactly why the A3B models are such good value.

Size ladder

• Qwen3.5-2B (2B dense, 256K ctx) — Q4 ≈ 3 GB → any 12 GB card runs it with huge headroom (details)
• Qwen3-4B-Instruct-2507 (4B dense, 256K ctx) — Q4 ≈ 4 GB → 12 GB; an easy daily driver for light tasks (details)
• Qwen3-4B-Thinking-2507 (4B dense, 256K ctx) — Q4 ≈ 4 GB → 12 GB; the reasoning-tuned sibling of the Instruct version (details)
• Qwen2.5-7B-Instruct (7.6B dense, 128K ctx) — Q4 ≈ 6 GB → 12 GB; an older classic that is still a community favorite (details)
• Qwen3-8B (8.2B dense, 128K ctx) — Q4 ≈ 6 GB → 12 GB; the most-downloaded Qwen in our catalog, with over 17 million downloads (details)
• Qwen3.5-9B (9B dense, 256K ctx) — Q4 ≈ 7 GB → 12 GB; the newest-generation pick at this size (details)
• Qwen2.5-14B-Instruct (14.8B dense, 128K ctx) — Q4 ≈ 10 GB → 12 GB, tight but workable (details)
• Qwen3-14B (14.8B dense, 128K ctx) — Q4 ≈ 10 GB → 12 GB; the biggest dense Qwen that still fits a 12 GB card (details)
• Qwen3-30B-A3B-Instruct-2507 (30.5B MoE, 3B active, 256K ctx) — Q4 ≈ 18 GB → 24 GB; MoE speed at a mid-range budget (details)
• Qwen3.5-35B-A3B (34.7B MoE, 3B active, 256K ctx) — Q4 ≈ 21 GB → 24 GB (details)
• Qwen3.5-27B (27B dense, 256K ctx) — Q4 ≈ 21 GB → 24 GB (details)
• Qwen3.6-35B-A3B (36B MoE, 3B active, 256K ctx) — Q4 ≈ 21 GB → 24 GB; the freshest A3B variant (details)
• Qwen3.6-27B (27.8B dense, 256K ctx) — Q4 ≈ 22 GB → 24 GB (details)
• Qwen2.5-32B-Instruct (32.8B dense, 128K ctx) — Q4 ≈ 26 GB → just over 24 GB, so plan on 48 GB (details)
• Qwen2.5-Coder-32B-Instruct (32.8B dense, 32K ctx) — Q4 ≈ 26 GB → 48 GB; a beloved coding specialist (details)
• Qwen3-Coder-Next (79.7B MoE, 3B active, 256K ctx) — Q4 ≈ 45 GB → 48 GB; huge coding quality, only ~3B active per token (details)
• Qwen3-Next-80B-A3B-Instruct (81.3B MoE, 3B active, 256K ctx) — Q4 ≈ 46 GB → 48 GB (details)
• Qwen2.5-72B-Instruct (72.7B dense, 128K ctx) — Q4 ≈ 56 GB → over 48 GB, so 96 GB territory (details)
• Qwen3.5-122B-A10B (122B MoE, 10B active, 256K ctx) — Q4 ≈ 70 GB → 96 GB (details)
• Qwen3-235B-A22B-Instruct-2507 (235B MoE, 22B active, 256K ctx) — Q4 ≈ 135 GB → 192 GB+ multi-GPU or unified-memory setups (details)
• Qwen3.5-397B-A17B (397B MoE, 17B active, 256K ctx) — Q4 ≈ 223 GB → 192 GB+ only (details)
• Qwen3-Coder-480B-A35B-Instruct (480B MoE, 35B active, 256K ctx) — Q4 ≈ 272 GB → the family flagship; workstation clusters only (details)

Family highlights

• MoE done right. The A3B / A10B / A22B naming tells you the active parameters per token. Because KV cache scales with active params in our formula, a 36B-A3B model needs roughly the same VRAM at 8K context as a dense 27B — while storing far more total knowledge in its experts.
• Long context is standard. Everything from the 2507 refresh onward offers 256K context; the older Qwen2.5 and first Qwen3 dense models sit at 128K. Remember that KV cache grows linearly with context, so budget extra VRAM if you plan to actually use those windows.
• One license, no drama. Every model here is Apache-2.0 — commercial use, fine-tuning and redistribution are all on the table.
• Specialist branches. Coder models target programming, Thinking variants favor long-form reasoning, and the Next line pushes the sparse-MoE architecture furthest.

Which one should you pick

• 12 GB (RTX 3060, RTX 5070 class): Qwen3.5-9B is the sweet spot; Qwen3-14B squeezes in at Q4 if you keep context modest. FP16 is realistic only for the 2B–4B models.
• 16 GB (RTX 4060 Ti 16GB, RTX 5060 Ti 16GB): same models as 12 GB, but with room to stretch context — a 14B at 16K context lands around 15 GB total.
• 24 GB (RTX 3090, RTX 4090 class): the A3B MoE trio (30B, 35B) or the dense 27B models — the best quality-per-dollar zone in the whole family.
• 48 GB (RTX 6000 Ada, A6000, dual 24 GB): dense 32B, or the 80B Next and Coder-Next MoEs.
• 96 GB (RTX PRO 6000, H100 class): Qwen2.5-72B dense or the 122B-A10B MoE.
• 192 GB+ (M2 Ultra 192, B200, multi-GPU): 235B, 397B and the 480B Coder.

Links

• Best GPUs for Qwen3
• Best models for 24 GB VRAM
• Best GPUs for coding models
• Guide: How much VRAM do local LLMs need?
• Guide: Running large models on consumer GPUs