Note: This article is adapted from the official Phala Network model listings on X. Check it here: https://x.com/phalanetwork/status/2058276251111174317
If you’ve been watching the AI infrastructure space, you already know that most inference providers make you pick a lane. You either get capable models or you get a platform you can actually trust with sensitive work. Phala Network is quietly making that trade-off irrelevant, and two new model listings are a good example of why.
What Just Got Listed
Two new models are now live on Phala’s inference platform.
1️⃣ The first is Qwen3.6 35B-A3B Uncensored, built on Alibaba’s latest mixture-of-experts architecture. It was fine-tuned by HauhauCS and quantized to FP8 precision by lamianlbe. In testing, it recorded zero refusals across 465 prompts, which tells you something about how it performs on real-world tasks that other models tend to sidestep.
Check it below ⬇️
🤗 https://huggingface.co/HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive
On Phala models 🔗: https://phala.com/models/phala/qwen3%2E6-35b-a3b-uncensored
2️⃣ The second is Gemma-4 26B Uncensored, referred to as the Heretic edition. It’s based on Google’s Gemma 4 MoE and was fine-tuned using the ARA method from Heretic v1.2.0 by coder3101, then quantized by cloud19. That fine-tuning brought its refusal rate from 100 out of 100 down to 11 out of 100, a significant shift for builders who need a model that actually completes the task.
Check it below ⬇️
🤗: https://huggingface.co/coder3101/gemma-4-26B-A4B-it-heretic
On Phala models 🔗: https://phala.com/models/phala/gemma-4-26b-a4b-uncensored
Why the Infrastructure Behind These Models Matters
The models themselves are notable, but what Phala brings to the table goes beyond the listings. Every model on the platform runs inside a TDX-attested H200 enclave, and responses are signed using ECDSA. In plain terms, that means each inference call comes with cryptographic proof that it ran in a secure, isolated environment and that the output was not altered between the model and your application.
For most use cases this might sound like extra detail. But if you’re running agent workflows, processing user data, or building anything where the integrity of the output matters, this is the kind of guarantee that changes how you architect things. You’re not just trusting the provider’s word. You have proof you can verify.
Why Phala Is Worth Following Right Now
Open, capable models have been available for a while. Secure compute environments have also existed for a while. What’s been missing is a platform that combines both without asking developers to manage the complexity themselves. Phala does that, and it keeps adding to the model catalog in a way that makes the platform more useful for production work over time.
Builders working on AI agents, institutions that need audit-ready inference, and developers who want flexibility without giving up reliability all have a reason to look at what Phala is building. The two models listed today are a good starting point, but the more important thing to understand is the infrastructure they’re running on.
Both models are available now at https://phala.com/models
Top comments (0)