vynaris/qwen3.6-35b-a3b-uncensored
Qwen3.6-35B-A3B-uncensored
The fast hosted option for high-volume authorized security testing, agents, and reduced-refusal research.
Input / MTok$1.00
Output / MTok$5.00
Context128K tokens
StatusHosted
Model directory + hosted API
Find an uncensored model—and know where you can actually use it.
Compare reduced-refusal model families, community variants, source repositories, practical tradeoffs, and hosted availability. Vynaris-hosted models are labelled separately from models listed for research and discovery.
Built for lawful, authorized red teaming, cyber defense, and model evaluation. Child sexual exploitation, human trafficking or exploitation, and non-consensual sexual content are prohibited. Only test systems and data you own or are authorized to assess.
Hosted now
Three model profiles. One API and one wallet.
These are the only directory entries currently exposed through public Vynaris model IDs. Select the exact ID in every OpenAI-compatible request.
API, variants and model card →
vynaris/qwen3.8-27b-uncensored
Qwen3.8-27B-uncensored
The balanced dense-model option for detailed authorized red-team analysis, coding, and long-context reviews.
Input / MTok$1.00
Output / MTok$7.00
Context128K tokens
StatusHosted
vynaris/deepseek-v4-flash-uncensored
DeepSeek-V4-Flash-uncensored
The highest-capability hosted tier for difficult multi-step security analysis and complex defensive engineering work.
Input / MTok$2.00
Output / MTok$11.00
Context128K tokens
StatusHosted
Across every size class
Explore uncensored model cards
Each page explains the family, capabilities, limitations, community variants, source repositories, use paths, and common questions. We add families based on community adoption and deployment interest.
Model guide9B class
Qwen3.5 9B Uncensored
A compact reduced-refusal Qwen family for local assistants, tool experiments, and fast evaluation on limited hardware.
- Architecture
- Dense transformer
- Variants linked
- 2
Model guide8B class
Qwen3-VL 8B Uncensored
A compact vision-language family for reduced-refusal image analysis and multimodal security evaluation.
- Architecture
- Vision-language transformer
- Variants linked
- 2
Model guide20B class
GPT-OSS 20B Uncensored
A reduced-refusal adaptation of the compact GPT-OSS family for reasoning experiments and authorized security research.
- Architecture
- Sparse mixture-of-experts
- Variants linked
- 1
Model guide18B–21B class
Llama 3.2 MoE Uncensored
Community mixture-of-experts builds combining Llama-derived experts for creative work, roleplay, and reduced-refusal evaluation.
- Architecture
- Community mixture-of-experts builds
- Variants linked
- 2
Model guide26B class
SuperGemma4 26B Uncensored
A community Gemma 4 adaptation targeting stronger instruction compliance with fewer refusals in a mid-size MoE profile.
- Architecture
- Sparse mixture-of-experts
- Variants linked
- 2
Model guide27B class
Qwen3.6 27B Uncensored
A popular dense Qwen family adapted by community publishers for creative work and reduced-refusal evaluation.
- Architecture
- Dense hybrid-attention transformer
- Variants linked
- 2
Model guide30B total / 3B active class
Qwen3 30B-A3B Uncensored
A compact-active-parameter Qwen MoE family that balances throughput, model breadth, and reduced-refusal experimentation.
- Architecture
- Sparse mixture-of-experts
- Variants linked
- 2
Model guideLarge sparse expert model
Gemma 4 E4B Uncensored
A popular community Gemma 4 adaptation pairing a large expert pool with sparse active computation and reduced refusals.
- Architecture
- Sparse mixture-of-experts
- Variants linked
- 1
How the directory works
Useful discovery without pretending every checkpoint is hosted.
One family page
Independent community releases are grouped around a readable model family instead of creating duplicate, thin pages.
Source attribution
Variant cards link to the publisher's Hugging Face repository, with visible publisher and license information.
Honest availability
Hosted badges appear only when an exact Vynaris model ID exists. Other pages show source and related-use options.
API use
Select a hosted model in every request.
Put the exact Vynaris model ID in the standard model field. No separate endpoint, provider configuration, or infrastructure detail is required.
curl https://api.vynaris.com/v1/chat/completions \
-H "Authorization: Bearer $VYNARIS_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "vynaris/qwen3.6-35b-a3b-uncensored",
"messages": [{"role": "user", "content": "Review this authorized test plan..."}],
"stream": true
}'Hosted pricing
Published rates, deducted from your normal balance.
Top-ups are $50, $100, $250, $500. A successful purchase of at least $50 in API credit qualifies the account for hosted models. Access remains available as the balance is spent down.
| Model | Input / 1M tokens | Output / 1M tokens | Context |
|---|---|---|---|
| Qwen3.6-35B-A3B-uncensored | $1.00 | $5.00 | 128K tokens |
| Qwen3.8-27B-uncensored | $1.00 | $7.00 | 128K tokens |
| DeepSeek-V4-Flash-uncensored | $2.00 | $11.00 | 128K tokens |
Retention
Hosted prompts and outputs are not persisted.
For hosted model IDs, Vynaris retains only the billing and operational metadata needed to charge correctly, run the service, investigate abuse reports, and answer lawful requests.
Cold starts and batching
Capacity sleeps when demand does.
The first request after an idle period can take longer while capacity starts. Concurrent requests queue into shared workers so warm capacity can batch demand.
FAQ
Questions about the directory and hosted API
Are all models in this directory hosted by Vynaris?
No. Only cards marked Hosted on Vynaris have a public Vynaris model ID. Directory-only cards describe community model families, link to their Hugging Face repositories, and identify a related hosted alternative without claiming the exact model is available.
What is an uncensored or reduced-refusal model?
It is generally a community-modified open-weight model intended to refuse fewer prompts. Uncensored, abliterated, heretic, and reduced-refusal are not standardized quality labels, so behavior must be evaluated per release.
Do you store prompts or model outputs?
For the Vynaris-hosted model IDs listed here, prompt and output bodies are not persisted. Vynaris retains operational metadata such as account, model, token counts, price, latency, status, and request ID.
Why can the first hosted request take longer?
Private capacity can sleep while idle. The first request may wait while the model service starts. Warm requests share the same workers, allowing concurrent demand to be queued and batched.
Is there a separate subscription or unlock?
No. Billing is unified. A successful purchase of at least $50 in API credit qualifies the account. Hosted access remains available as that balance is spent down.
Does reduced refusal make a model accurate or safe?
No. Reduced refusal can also reduce caution and calibration. Validate important outputs, follow each repository's license, and use models only for lawful, authorized work.
Independent directory, explicit sources.
Model names belong to their respective projects and publishers. Vynaris does not imply affiliation or endorsement, and the directory does not mirror model files. 3 entries are currently hosted; the rest are research guides.