Models
ReadyGPU TEE

Phala: Gemma-4 26B-A4B Uncensored (Heretic)

Model IDphala/gemma-4-26b-a4b-uncensored

Uncensored "Heretic" variant of google/gemma-4-26B-A4B-it created using Heretic v1.2.0 with the Arbitrary-Rank Ablation (ARA) method and row-norm preservation. Refusals drop from 100/100 to 11/100 with KL divergence 0.0499 vs the base model. The base Gemma 4 26B A4B is a Mixture-of-Experts model with 25.2B total / 3.8B active parameters (8 active / 128 total experts), 30-layer transformer with hybrid local sliding (1024) + global attention, supporting a 256K context window. Natively multimodal (text + images, variable aspect ratios). Strong on coding, reasoning, function calling, with native system prompt support across 35+ languages. Served on Phala in TDX-attested H200 enclave with end-to-end ECDSA response signing; vLLM-compatible FP8-Static quantization by cloud19 (router excluded from quantization).

input

$0.15/M

output

$0.70/M

cache read

$0.05/M

context

262K

added

May 23, 2026

Supported API shape

input

text · image

output

text

tools

Supported

json mode

Supported

Verification

Available evidence. Use a verified client to authenticate the quote, connection, and receipt.

Verify private inference

receipt

signed receipt + hashes

attestation

gateway report

session

attested upstream

serving provider

read from receipt session

Provider

Phala

GPU TEE

input

$0.15/M

output

$0.70/M

cache read

$0.05/M

context

262K

Performance comparison

Last 72h · UTC

More models

Other private inference routes.

View catalog
encrypted

Phala: DeepSeek V4.1 Flash 1017 Uncensored

Uncensored variant of DeepSeek V4.1 Flash 1017, built as a lightweight LoRA adapter served unmerged on the official FP8 weights. Refuses far less while keeping the base model's capabilities, tool calling, image input and 1M-token context. Served with SGLang in a TDX-attested enclave with NVIDIA confidential computing. Reasoning is on by default and can be turned off or capped per request.

context

1.0M

input

$0.30/M

$0.006/M cache read

encrypted

Qwen: Qwen3 ASR 0.6B

Qwen3-ASR 0.6B is the smaller, lower-cost open-weight speech recognition model in Alibaba's Qwen3-ASR family, transcribing speech in 30 languages and 22 Chinese dialects with automatic language detection. Served on Phala GPU TEE. It returns plain text for audio up to 10 minutes and does not return timestamps.

price

$0.0002 / min

encrypted

Qwen: Qwen3 ASR 1.7B

Qwen3-ASR 1.7B is the larger open-weight speech recognition model in Alibaba's Qwen3-ASR family, transcribing speech in 30 languages and 22 Chinese dialects with automatic language detection. Served on Phala GPU TEE. It returns plain text for audio up to 10 minutes and does not return timestamps.

price

$0.00045 / min

encrypted

MoonshotAI: Kimi K3.1 Pro

Kimi K3.1 Pro is Moonshot AI's flagship open-weight multimodal reasoning model and a post-trained successor to Kimi K3, with a 1M-token context window. It is built for long-horizon coding, 3D creation, financial research, and end-to-end agentic knowledge work, with native image and video understanding.

context

1.0M

input

$2.00/M

$0.20/M cache read