1× Solo
DGX Spark
Qwen3.8-Flash-NextSolo load79.5- Qwen3.8-27B27B, SGLang72
Ling-3.0-Flash
GLM 5.3 FlashEXL3 K2, one box18.6
Open weights. Uncensored models.
Private inference.
Open to everyone.
Choose your model and where it runs. Run locally to keep prompts on your device, or use independent hosts on the network.
Hosts keep 80% of each completed job. You choose your rate.
No chat logs. Prompts are erased when a host picks them up. Replies are erased 10 minutes after they finish. We keep only token counts for billing. How privacy works
Live now
All models & pricesChecking availability…
Rates apply to input and output tokens. Speed is desk-measured where marked, otherwise host-reported; tests do not verify model weights.
Your model. Your call.
Text inference on the network. Image, voice, and audio models to explore locally.
- Text & codeOn the network
- ImagesExplore locally
- VoiceExplore locally
- Speech-to-textExplore locally
- AudioExplore locally
TRENDING
- 1
DeepSeek v4.1 FlashHOT - 2Empero Qwen3.8-35B-A3BHOT
- 3
Gemma-4 26B UncensoredHOT - 4Gemma-4 26B-A4B-itHOT
- 5GLM 5.3 FlashHOT
STREET · DGX SPARK
Sample data · 2026-11-04
$8,050▼ 6.7% 7d
TAPE
@fillagrew “351 tokens/sec on Qwen3.5 2B at Q4_K_M”
@Youssofal_ “Peak Speed: 126.5 TPS.”
Street stacks
since Sep 20
2× Dual
DGX Spark
- GLM 5.3 FlashGLM on both60.4
- GLM 5.3 FlashTensorFold, 4 streams60.4
- DeepSeek v4.1 Flash103
- Qwen3.8-Flash-Next45
MiMo-V2.6-Flash
3× Triple
DGX Spark
- GLM 5.3 FlashAll three
- GLM 5.3 Flash · Qwen3.8-Flash-NextGLM orch on 2×, Qwen worker on 1×
- DeepSeek v4.1 FlashHomogeneous
- GLM-5.2NVFP4
4× Quad
DGX Spark
- GLM 5.3 Flash · Qwen3.8-Flash-NextGLM orch 2× + Qwen worker 2×104
- GLM 5.3 FlashHomogeneous104
- DeepSeek v4.1 FlashHomogeneous
6× 6+
DGX Spark
- GLM 5.3 Flash · Qwen3.8-Flash-Next · DeepSeek v4.1 FlashGLM orch + Qwen worker + DeepSeek scanner
Uncensored32
Every liberated weight. A number is a quoted tok/s.
Quoted
No quote yet
Payback ranked
US price. Each model at its suggested price, host keeps 80%. 20% duty. Left is faster.
| # | Quoted pairs | Months | ROI |
|---|---|---|---|
| 1 | Qwen3.8-27B · RTX 5090 | 10 mo | 117% |
| 2 | Qwen3.8-27B · RTX 5090 | 15 mo | 80% |
| 3 | Gemma-4 26B-A4B-it · RTX 5090 32GB | 17 mo | 71% |
| 4 | Qwen3.8-27B · RTX 5090 | 21 mo | 57% |
| 5 | Qwen3.8-27B · RTX 5060 Ti 16GB | 2.1 yr | 47% |
| 6 | Qwen3.8-27B · RTX 5060 Ti 16GB | 2.7 yr | 37% |
| 7 | Ternary Bonsai 2 27B · RTX 5060 Ti 16GB | 2.8 yr | 36% |
| 8 | Qwen3.5 2B · RTX 5090 32GB | 2.9 yr | 34% |
Articles
2026-09-28How to bench a cyber weightA cyber bench ranks a published decode, with the publisher's name on the row. The first shelf is /cyber. This desk did not remeasure it.
2026-09-28A shell is not a weightMikko Haapoja posted one line and a link. The link is NVIDIA OpenShell. It keeps an agent inside a policy. It does not serve a model, and this desk did not run it.
2026-09-24Which box pays backStreet price comes back only if the tokens, at $1 per million, beat the watts. A Spark bought to resell that rate does not.
2026-09-23Inference, not a JEPAA decode tok/s figure is a serving number. Training a JEPA, or drawing a picture, wants a different machine.
2026-09-22The interface is the argumentThe home rank, the pool shelf, and the pair page are the argument. The desk shows those. It does not show the chain.