8 open-weight models — Qwen, Llama, DeepSeek, Mistral, Gemma and more — all served as 2-bit (Q2_K) compressed weights. Prices are per 1M tokens. Call any of them at /v1/chat/completions with your API key.
Quality: independent benchmarks (Unsloth Dynamic 2.0, 5-shot MMLU) show 2-bit Q2_K_XL retains ≈96% of full-precision accuracy (e.g. Gemma-3-27B: 68.7% vs 71.5%), with perplexity within ~2%. A deliberate cost/quality trade-off — great for high-volume and cost-sensitive workloads.
Tiny · ≤2B1
Model
Input
Output
Tier
qwen3-1.7b-q2
$0.20
$0.60
L4 · standard
Small · 3–9B2
Model
Input
Output
Tier
gemma-3-4b-vision
$1.00
$3.00
L4 · standard
qwen3-8b-q2
$0.50
$1.50
L4 · standard
Mid · 10–20B2
Model
Input
Output
Tier
deepseek-r1-distill-qwen-14b-q2
$1.00
$3.00
L4 · standard
qwen3-14b-q2
$1.00
$3.00
L4 · standard
Large · 22–40B3
Model
Input
Output
Tier
gemma-3-27b-q2
$1.60
$4.80
L4 · standard
qwen3-coder-30b-a3b-q2
$1.60
$4.80
L4 · standard
qwen3-32b-q2
$1.60
$4.80
L4 · standard
All models are 2-bit (Q2_K-family) quantized — a deliberate cost/quality trade-off, not the vendors' full-precision endpoints. The live list is always at GET /v1/models. See the docs to start.