Qwen3 8B
Data checked: 2026-10-12. Review interval shown below. Prices need review after 30 days; specifications and benchmark collections after 90 days. Dates describe the recorded check, not a live retailer quote.
- Verified Checked against the primary source (for example the model’s config.json on Hugging Face).
- Sourced Taken from a cited source; not yet independently checked.
- Estimate Calculated with the published formulas; not a measurement.
- Sample Placeholder data for layout; not a real figure.
- Not yet measured No public, citable score found for this benchmark yet. We never fill the gap with a guess.
Verified metadata is not a measured performance result. Only cited runs matching model, device, quant and recorded context can replace speed estimates. Estimates assume one stream and F16 KV cache, with a ±25% uncertainty band. A published run applies only to its recorded runtime and conditions; missing settings are unknown.
Evidence labels · Freshness policy · Calculation methodology
At a glance
A small dense model with switchable thinking. The pick for 8 to 12 GB cards when you want Apache-2.0 terms.
At Q4_K_M and 8k context it needs about 6.8 GB: 5.0 GB of weights, 1.2 GB of KV cache and 0.6 GB of runtime overhead. It is dense, so every token reads all 5.0 GB of weights; speed scales with memory bandwidth.
Estimated setup
Quant | Q4_K_M |
|---|---|
Smallest standard card size at 8k | 16 GB |
Max context, 24 GB card | 32k |
Runtimes | llama.cpp, Ollama, LM Studio, MLX |
Quantization options
Total memory = weights + KV cache (F16) + runtime overhead. File sizes marked Verified are the sum of GGUF files from Qwen/Qwen3-8B-GGUF on the Hugging Face API. KV cache costs 0.147 MB per token before sliding-window savings.
Quant | bpw | File | At 8k | At 32k | At 128k | KL-divergence | 24 GB card, 8k |
|---|---|---|---|---|---|---|---|
| Q3_K_M | 3.91 | 4.0 GBEstimate | 5.8 GB | 9.6 GB | — | 0.071Sample | Fits |
| Q4_K_M | 4.85 | 5.0 GBVerified | 6.8 GB | 10.6 GB | — | 0.017Sample | Fits |
| Q5_K_M | 5.69 | 5.9 GBVerified | 7.6 GB | 11.4 GB | — | 0.008Sample | Fits |
| Q6_K | 6.56 | 6.7 GBVerified | 8.5 GB | 12.3 GB | — | 0.004Sample | Fits |
| Q8_0 | 8.50 | 8.7 GBVerified | 10.5 GB | 14.3 GB | — | 0.001Sample | Fits |
| F16 | 16.00 | 16.4 GBEstimate | 18.2 GB | 22.0 GB | — | reference | Fits |
Why this matters: What is quantization? · What is a KV cache?
Decode speed by device
Q4_K_M, single stream, 8k context. Estimates use the brief §7.3 formula (η × bandwidth ÷ bytes per token) and are ±25% until measured.
Speed table
Device | Verdict | Decode, tok/s | Range | Runtime |
|---|---|---|---|---|
| GeForce RTX 5090 32 GB | Fits | 172.4Estimate | ±25% | llama.cpp (estimate) |
| GeForce RTX 4090 24 GB | Fits | 97.0Estimate | ±25% | llama.cpp (estimate) |
| Radeon RX 7900 XTX 24 GB | Fits | 92.3Estimate | ±25% | llama.cpp (estimate) |
| GeForce RTX 3090 24 GB (used) | Fits | 90.0Estimate | ±25% | llama.cpp (estimate) |
| 2× RTX 3090 (used) 48 GB | Fits | 90.0Estimate | ±25% | llama.cpp (estimate) |
| Mac Studio M5 Ultra (base) | Fits | 86.6Estimate | ±25% | llama.cpp (estimate) |
| Mac Studio M5 Ultra · 512 GB | Fits | 86.6Estimate | ±25% | llama.cpp (estimate) |
| GeForce RTX 5060 Ti 16 GB | Fits | 43.1Estimate | ±25% | llama.cpp (estimate) |
| MacBook Pro M4 Max 40c · 128 GB | Fits | 39.4Estimate | ±25% | llama.cpp (estimate) |
| Mac Studio M5 Max (base) | Fits | 33.2Estimate | ±25% | llama.cpp (estimate) |
| GeForce RTX 4060 Ti 16 GB | Fits | 27.7Estimate | ±25% | llama.cpp (estimate) |
| NVIDIA DGX Spark FE 128 GB | Fits | 19.7Estimate | ±25% | llama.cpp (estimate) |
| Ryzen AI Max+ 395 mini PC 128 GB | Fits | 18.5Estimate | ±25% | llama.cpp (estimate) |
| CPU only · Ryzen 9 9950X + 128 GB DDR5 | Fits | 7.2Estimate | ±25% | llama.cpp (estimate) |
Compare with similar-sized models
Model | Parameters | Quality score | Licence |
|---|---|---|---|
| Qwen3 8B | 8.2 B | Not measuredSample | Apache-2.0 |
| Llama 3.1 8B Instruct | 8.0 B | 1,187Sourced | Llama 3.1 Community License |
Placeholder: this model has no entry in the LMArena text leaderboard dataset (checked 2026-10-12). Not a real score.
llama-perplexity --kl-divergence on a fixed corpus. Speed and quality are never merged into one score. Sample values are illustrative and must not be used to select a model.Architecture
Parameters, total | 8.19 BVerifiedsafetensors.total from the HF API. Source |
|---|---|
Parameters, active per token | 8.19 BVerifiedDense: all parameters active. Source |
Layers | 36VerifiedSource |
Query heads | 32VerifiedSource |
KV heads (GQA) | 8VerifiedSource |
Head dimension | 128VerifiedSource |
Hidden size | 4096VerifiedSource |
Operating conditions
Max context | 32,768 tokensSourcedconfig.json lists max_position_embeddings 40,960; the model card gives 32,768 native and 131,072 with YaRN. Source |
|---|---|
Licence | Apache-2.0VerifiedSource |
Weights | Qwen/Qwen3-8B |
GGUF files | Qwen/Qwen3-8B-GGUF |
Usable memory, discrete | VRAM − 0.5 GB2 |
Usable memory, unified | RAM × 0.752 |
Fits when | total ≤ 90% of usable2 |
1 Estimates: decode ≈ η × bandwidth ÷ (active weights + KV cache), η 0.60 discrete / 0.45 unified / 0.50 CPU (concept brief §7.3). Not measurements.
2 Fit rules from the concept brief §7.3, including the 0.75 unified-memory factor used by canirun.ai.
Flags: Verified = checked against config.json / HF API; Sourced = from a cited source; Estimate = formula; Sample = placeholder. Markdown · JSON · Report a problem