ReadyGPU TEE

Phala: Qwen3.8 27B Uncensored (Aggressive)

Model IDphala/qwen3.8-27b-uncensored

HauhauCS Qwen3.8-27B Uncensored Aggressive, using the pinned Q8_K_P GGUF and BF16 vision projector. Served with SGLang on a Phala TDX-attested H200, with text and image input and a 262144-token context.

Start building Docs

input

$0.30/M

output

$1.50/M

cache read

$0.05/M

context

262K

created

Sep 16, 2026

Supported API shape

input

text · image

output

text

tools

Supported

json mode

Supported

Verification

receipt

x-receipt-id

attestation

gateway report

session

attested upstream

provider

Phala

Provider

Phala

GPU TEE

input

$0.30/M

output

$1.50/M

cache read

$0.05/M

context

262K

Prompt caching

Providers can reuse a repeated prompt prefix automatically. No extra request parameter is needed. Put stable system instructions and reference material first, then add the changing question at the end. Cache availability and minimum prefix length depend on the provider and model.

Cached input tokens use the listed cache-read rate. When a model has no cache-read rate, the normal input rate applies. Output tokens use the output rate. A listed rate does not guarantee a cache hit.

Check usage.prompt_tokens_details.cached_tokens in the response. Cached tokens are part of prompt_tokens, so subtract them before calculating uncached input cost. In this example, 1,792 input tokens are cached and 113 are uncached.

"usage": {
  "prompt_tokens": 1905,
  "prompt_tokens_details": { "cached_tokens": 1792 },
  "completion_tokens": 64
}

Performance comparison

Last 72h · UTC

$ pip install openai

1from openai import OpenAI
2import os
3
4MODEL = "phala/qwen3.8-27b-uncensored"
5
6client = OpenAI(
7    base_url="https://inference.phala.com/v1",
8    api_key=os.environ["PHALA_API_KEY"],
9)
10
11response = client.chat.completions.create(
12    model=MODEL,
13    messages=[{"role": "user", "content": "Run this privately."}],
14    extra_body={"provider": {"aci_verified": True}},
15)
16
17print(response.id)
18print(response.choices[0].message.content)

$ npm install openai

1import OpenAI from "openai"
2
3const MODEL = "phala/qwen3.8-27b-uncensored"
4
5const openai = new OpenAI({
6  baseURL: "https://inference.phala.com/v1",
7  apiKey: process.env.PHALA_API_KEY,
8})
9
10const response = await openai.chat.completions.create({
11  model: MODEL,
12  messages: [{ role: "user", content: "Run this privately." }],
13  provider: { aci_verified: true },
14})
15
16console.log(response.id)
17console.log(response.choices[0].message.content)

1curl https://inference.phala.com/v1/chat/completions \
2  -H "Authorization: Bearer $PHALA_API_KEY" \
3  -H "Content-Type: application/json" \
4  -d '{
5    "model": "phala/qwen3.8-27b-uncensored",
6    "messages": [
7      { "role": "user", "content": "Run this privately." }
8    ],
9    "provider": {"aci_verified": true}
10  }'

More models

Other private inference routes.

View catalog

encrypted

Jared Palmer: Kev 4B

Kev 4B is an open-weight System One model from Jared Palmer, fine-tuned from Qwen3.5 4B Base. It makes fast, structured decisions for software: given a state and typed questions, it returns calibrated answers instead of text, as a yes/no probability (noul), a pick among options (choice), or a level on a scale (score). Call it through POST /v1/systemone, compatible with the TypeSafe SDK; it does not serve chat. Served on Phala in a TDX-attested enclave with an 8,192-token context.

context

input

$0.042/M

encrypted

Qwen: Qwen3.8 27B

Qwen3.8 27B is an open-weight dense vision-language model from Qwen. It is suited for coding, professional workflows, research, multimodal interaction, and long-running agent tasks, with flexible thinking that can be enabled or disabled. Served on Phala in a TDX-attested enclave.

context

input

$0.20/M

$0.05/M cache read

encrypted

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

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).

context

262K

input

$0.15/M

$0.05/M cache read

encrypted

Sentence Transformers: all-MiniLM-L6-v2

The all-MiniLM-L6-v2 embedding model maps sentences and short paragraphs into a 384-dimensional dense vector space, enabling high-quality semantic representations that are ideal for downstream tasks such as information retrieval, clustering, similarity scoring, and text ranking.

context

512

input

$0.005/M