AxForge offering · Request
A mixture-of-experts Qwen with roughly 3B active parameters per token, validated on AxForge hardware with a dedicated DGX Spark deployment option for your traffic only, subject to confirmed availability, configuration and licence. An OpenAI-compatible endpoint on your own machine, operated by AxForge in Málaga, Spain.
Currently request-only. Add your request to the public forum; this is not a booking or a promised launch.
Why AxForge
| MoE built for GB10 | 35B total, roughly 3B active parameters per token, so single-stream decode is faster than a dense model of similar size. On the community SparkBench leaderboard (PBM, 4k context, single GB10, NVFP4 + MTP) it measures 86.3 tokens/s. Labelled community data, not ours. |
|---|---|
| Your machine, your endpoint | For an agreed deployment, subject to availability: A dedicated NVIDIA DGX Spark (GB10, 128 GB unified memory) reserved for you, running qwen3.6-35b-a3b-nvfp4 behind an OpenAI-compatible /v1 that serves only your traffic. |
| EU-hosted, zero prompt retention | For an agreed deployment, subject to availability: Hosted in Málaga, Spain (eu-es-1). Prompts and completions are processed in memory, not logged, not retained, never used to train. It is the same policy as the serverless API. |
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, NVFP4 |
| Context window | 65,536 tokens |
| Hardware | NVIDIA DGX Spark (GB10, 128 GB unified memory), owned and operated by AxForge |
| Rental term | Hour, week, month or year |
| Hardware pricing | €1.19/hour on demand · €0.95/hour by the week · €0.91/hour by the month · €0.85/hour by the year |
| Managed service | Quoted per deployment |
| Region | Málaga, Spain (eu-es-1) |
Qwen3.6 35B A3B catalogue reference: specifications and requirements. List prices in EUR.
How it works
| 1 | Add a public request describing your workload, context needs and preferred rental term. A request does not reserve hardware or promise a deployment. |
|---|---|
| 2 | If AxForge can offer a deployment, agree the configuration, hardware rental, licence and managed-service price in writing before anything is billed. |
| 3 | Only after availability and terms are confirmed would AxForge provision the agreed dedicated system and model build. |
| 4 | After provisioning, point your OpenAI SDK at your own endpoint with model qwen3.6-35b-a3b-nvfp4. |
| 5 | After provisioning, adjust the term (hour, week, month or year) as your workload settles. |
Request model or sign in to start.
FAQ
This page describes the managed deployment option. Availability, the model build, hardware, licence and price must be agreed before a deployment can be offered. A public request does not provision or reserve a machine. Check the console for the current serverless offering list.
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.
On the community SparkBench leaderboard (PBM methodology, 4k context, single GB10), this model with NVFP4 + MTP measures 86.3 tokens/s single-stream. Labelled community data. AxForge has not published its own tuned numbers yet.
For a deployment agreed after availability is confirmed: A dedicated NVIDIA DGX Spark (GB10, 128 GB unified memory) rented by the hour, week, month or year, running Qwen3.6 35B A3B behind an OpenAI-compatible endpoint on your own machine, hosted in the EU with zero prompt retention.
For a deployment agreed after availability is confirmed: Two parts: the DGX Spark hardware rental (by the hour, week, month or year, with longer terms earning the lower rate) and the managed service, quoted per deployment. Both are confirmed in writing before anything is billed.
Check the console for current Qwen3.8 27B serverless availability. Launch pricing per million input and output tokens is on its page.