Qwen3-VL-2B-Thinking-FP8
Qwen3-VL-2B-Thinking-FP8 has 2.44B parameters (1.72B active per token) and takes 3.47 GB on disk. See where its parameters live, layer by layer, and what hardware it needs.
At a glance
| Qwen/Qwen3-VL-2B-Thinking-FP8 | |
|---|---|
| Repo created | Oct 20, 2025updated Nov 26, 2025 |
| Model type | DenseFull attention (GQA) |
| Inputs | Text + images407M vision encoder |
| Total parameters | 2.44B |
| Active per token | 1.72B71% of the model |
| Experts | None (dense) |
| Max context (from config) | 256K tokens |
| Layers | 28 |
| On disk | 3.47 GB1 file |
| Precision | BF16 (59%), FP8 E4M3 (41%) |
| Quantization | FP8 E4M3blocks of 128×128 · 58% of parameters · embeddings & output head and vision encoder kept in BF16 |
| Fewest GPUs | 1× RTX 4060a single 8 GB card · weights only |
| Family | Original |
| License | apache-2.0 |
| GitHub | QwenLM/Qwen3-VL |
Where the parameters live
Each bar is one model, split by component and drawn on the same scale.
Anatomy map
Every rectangle is one component in one layer, with area proportional to its size. When you compare two models, both maps use the same scale, so a model twice the size covers twice the area.
Layer map
The model's layers from input (left) to output (right): which kind of attention each one uses, and whether its feed-forward block is a mixture of experts or a single dense network.
One token's route
In a mixture-of-experts layer, a router picks a few experts for each token. This shows how many that is, drawn to scale: the lit cells are one token's share of a single layer, and everything else sits idle in memory. Which cells light up is a random example, not the model's real router.
Can my GPU handle this model?
At a datacenter
The fewest GPUs of each type whose combined memory can hold the weights, assuming 90% of each GPU's memory is usable. Real deployments need more memory for context (KV cache), and usually round up to 8 or 16 GPUs so the model splits evenly. The list covers NVIDIA data-center GPUs widely rentable as of September 2026, best first.
At home
Whether the weights fit on popular home GPUs, assuming 90% of the card's memory is usable. People at home usually run 4-bit versions (GGUF files for llama.cpp, Ollama or LM Studio), which are much smaller than what labs publish, so both are shown. The 4-bit size is an estimate of about 4.5 bits per parameter. "Tight" means the weights fit but leave little room for context.
Context memory
Besides its weights, a model needs memory for the text it's working on (the "KV cache"), and it grows with every token of context. Figures are for one conversation stored in 16-bit precision: each simultaneous conversation needs its own, and servers that store it in 8-bit (FP8) need half as much.