Qwen 3.6 27B FP8
Alibaba's 27B open-weight coding specialist with hybrid DeltaNet attention, agentic reasoning, and near-lossless FP8 quantization.
Overview
What is Qwen 3.6 27B FP8
Qwen 3.6 27B FP8 is Alibaba's open-weight, 27-billion-parameter dense language model released in April 2026. It combines Gated DeltaNet and attention layers for efficient long-context inference, and is optimized for agentic coding, reasoning, and tool use. The FP8 quantization preserves near-original accuracy while reducing memory footprint.
Running it privately on Venice
On Venice, Qwen 3.6 27B FP8 runs inside a TEE with end-to-end encryption and zero retention — your prompts are never stored or profiled. You get open-weight sovereignty plus native tool use, reasoning, web search, and code-optimized output, all without Big-Tech surveillance.
Assessment
Strengths and limitations
- Open-source Apache 2.0 weights: permissionless to self-host, audit, and fine-tune without vendor lock-in.
- Hybrid Gated DeltaNet + Attention architecture delivers efficient inference and long-context processing.
- Strong agentic coding and repository-level reasoning with native tool use and function calling on Venice.
- Supports reasoning mode and web search, making it effective for research and iterative development.
- FP8 compression preserves near-original accuracy while significantly lowering VRAM requirements versus full-precision inference.
- Not uncensored: the model retains safety alignment and may refuse certain prompts.
- Dense 27B size still requires a modern GPU with substantial VRAM; smaller quants or MoE alternatives fit tighter budgets.
- Venice hosts the text endpoint, so the base model's multimodal vision and video features are not exposed.
- European language coverage and conservative refusal patterns may lag behind some rivals in specific locales.
Capabilities
What it supports
- Tool use / function calling
- Vision (image input)
- Reasoning
- Web search
- Code-optimized
- Structured output (JSON schema)
- Audio input
- Video input
- Multiple image inputs
- Log probabilities
Specifications
Datasheet
- Maker
- Alibaba Cloud (Qwen team)
- Released
- April 22, 2026
- Architecture
- Dense hybrid (Gated DeltaNet + Gated Attention), 64 layers, 5120 hidden dim
- Parameters
- 27B
- Open weights
- Yes — Apache 2.0
- Quantization
- FP8 (fine-grained, block size 128)
- Context window
- 256K tokens
- Max output
- 32.768K tokens
- Capabilities
- Function calling, Reasoning, Web search, Code-optimized
- Privacy on Venice
- Private — zero retention
- Available on Venice since
- Jul 2026
API
Call it from your code
Venice exposes an OpenAI-compatible API. Point your base URL at Venice and pass the model id.
curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "e2ee-qwen3-6-27b",
"messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
}'Pricing
What it costs on Venice
Billed per token on Venice: $0.35 per 1M input tokens and $3.46 per 1M output tokens.
New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.
Alternatives
How it compares
| Model | Context window | Strongest at | Open weights | Price (Venice) |
|---|---|---|---|---|
| Qwen 3.6 27B FP8 | 256K tokens | Agentic coding & reasoning | Yes | $0.35 in · $3.46 out / 1M |
| DeepSeek V3.2 | 160K tokens | General reasoning & MoE efficiency | Yes | $0.33 in · $0.48 out / 1M |
| Kimi K2.6 | 256K tokens | Long-context document analysis | Yes | $0.75 in · $3.50 out / 1M |
| Google Gemma 4 31B Instruct | 256K tokens | Lightweight multilingual & safety | Yes | $0.12 in · $0.36 out / 1M |
| Claude Sonnet 4.6 | 1M tokens | Enterprise instruction following | No | $3.60 in · $18 out / 1M |
Open-weight dense model with hybrid DeltaNet attention and near-lossless FP8 quantization.
Use cases
What it is good for
- 01Agentic software engineering — repository-level reasoning, frontend workflows, and iterative debugging.
- 02AI coding assistants with tool use — function calling to run tests, query APIs, or search the web.
- 03Long-document analysis and summarization leveraging the 256K token context window.
- 04Private research and writing with web search and thinking preservation enabled.
- 05Self-hosted deployment via HuggingFace, vLLM, or SGLang for full data sovereignty.
Prompting
Getting better results
Enable reasoning mode for complex coding or math problems to unlock step-by-step thinking chains.
Use function calling to let the model execute commands, search the web, or interact with external tools within a conversation.
For long-context tasks, explicitly reference earlier sections of the document to improve recall across the 256K window.
When self-hosting the FP8 weights, use an inference engine that supports fine-grained FP8 (block size 128) for best fidelity.
Version history
Preceding open-weight Qwen generation.
Current dense flagship with hybrid attention and FP8 release.
MoE sibling with 35B total / 3B active parameters.
FAQ
Frequently asked questions
Qwen 3.6 27B FP8 is Alibaba's open-weight, 27-billion-parameter dense language model released in April 2026. It uses a hybrid Gated DeltaNet and attention architecture, supports tool use and reasoning, and is distributed under the Apache 2.0 license.
Venice bills per token at $0.35 per 1M input tokens and $3.46 per 1M output tokens. Cached input is $0.17 per 1M tokens. There is no subscription required.
Yes. The model is released under Apache 2.0 with open weights available on HuggingFace. You can self-host it with Transformers, vLLM, SGLang, or KTransformers.
Yes. On Venice it supports function calling, reasoning, web search, and code-optimized generation. You can use these capabilities through the Venice chat interface or API.
No. The model retains its native safety alignment and may refuse certain requests. Venice runs it privately with zero retention, but does not alter its underlying refusal behavior.
Choose Qwen 3.6 27B FP8 for agentic coding, longer context, and repository-level reasoning. Choose DeepSeek V3.2 for lower per-token pricing and strong general reasoning, especially if 160K context is sufficient.
Yes. The FP8 weights are published on HuggingFace and are compatible with HuggingFace Transformers, vLLM, SGLang, and KTransformers. The quantization uses fine-grained FP8 with block size 128.
It runs under Venice's private tier with zero retention, inside a TEE with end-to-end encryption. Your prompts are not stored, profiled, or used for training.
Run Qwen 3.6 27B FP8 privately
No prompt logging. No data used for training.