Dedicated model · deployable on request
A mixture-of-experts Qwen with roughly 3B active parameters per token, validated on AxForge hardware and deployed on a dedicated DGX Spark for your traffic only. Not on the shared API today.
Specifications
| Model | Qwen3.6 35B A3B — mixture-of-experts, latest Qwen generation |
|---|---|
| Served model name | qwen3.6-35b-a3b-nvfp4 |
| Architecture | MoE — 35B total, roughly 3B active parameters per token |
| Context window | 65,536 tokens |
| Quantization | NVFP4 |
| Availability | Deployable on request — not on the shared API today |
| Hardware | NVIDIA DGX Spark (GB10, 128 GB unified memory) — owned and operated by AxForge |
| Region | eu-se-1 · Stockholm, Sweden |
| Pricing | On request — talk to an engineer |
Performance
| Metric | Value | Condition |
|---|---|---|
| Decode speed | 30.7 tokens/s | single stream, MoE with ~3B active parameters |
Measured on our production DGX Spark node, single-stream, 2026-08. Faster single-stream than the dense Qwen3.8 27B we serve (15.9 tokens/s measured) because only ~3B parameters are active per token. For quality benchmarks, see the official model card — we publish only numbers we measured ourselves.
Deployment
| System | A dedicated NVIDIA DGX Spark (GB10, 128 GB unified memory), reserved for you |
|---|---|
| Commercial model | Monthly rental |
| Endpoint | OpenAI-compatible /v1 on your own machine — only your traffic |
| Location | Hosted in the EU (eu-se-1 · Stockholm, Sweden) |
| Data handling | Zero prompt retention, same policy as the shared API |
The model is validated on AxForge hardware and brought hot on your system. After deployment — and only then — your endpoint speaks the OpenAI API:
# After your dedicated node is live. This does not work on api.axforge.ai today. from openai import OpenAI client = OpenAI(base_url="https://<your-endpoint>/v1", api_key="<your-key>") r = client.chat.completions.create( model="qwen3.6-35b-a3b-nvfp4", messages=[{"role": "user", "content": "Hello from your own machine"}], )
Data & privacy
Prompts and completions are processed in memory in Sweden — not written to disk, not logged, not retained, and never used to train anything. We keep only request metadata (token counts, timestamps, status) for billing and operations. The full policy is at axforge.ai/privacy.
FAQ
No. It is not on the shared API today. It is deployable on request: validated on AxForge hardware and brought hot on a dedicated DGX Spark for your traffic only. The shared API serves Qwen3.8 27B.
Qwen3.6 35B A3B is a mixture-of-experts model: 35B total parameters, roughly 3B active per token. That is why its single-stream decode is faster than a dense model of similar size.
We measured 30.7 tokens/s single-stream on our production DGX Spark node, 2026-08. We publish measured numbers only.
A dedicated NVIDIA DGX Spark (GB10, 128 GB unified memory) rented monthly, running Qwen3.6 35B A3B behind an OpenAI-compatible endpoint on your own machine, hosted in the EU with zero prompt retention.
Pricing on request — talk to an engineer and we will quote the monthly rental.
Qwen3.8 27B is served now on the shared API at €0.29 per 1M input tokens and €1.77 per 1M output tokens (launch pricing).