Qwen3.8-27B-FP8

Qwen3.8-27B-FP8 brings the Qwen3.8 family’s agentic and multimodal advances into a deployment-friendly 27B dense model by using fine-grained FP8 post-training quantization. The result is a compact artifact that preserves most of the full model’s capabilities—vision, video, long-horizon reasoning and multi-step agent execution—while being practical to serve with modern inference stacks.

Key Capabilities

• Architecture and quantization: 27B parameters, 64 layers, 5120 hidden dim, and FP8 fine-grained post-training quantization with block size 128 — so you get a smaller memory footprint while retaining performance close to the original dense model.
• Native vision and video: built-in vision encoder and video support, enabling image-to-text and hour-scale video understanding for visual math, chart reading, document QA and more.
• Ultra-long context and agent control: natively supports 262,144 tokens (extensible toward 1,000,000 using YaRN/RoPE scaling). Thinking-mode is enabled by default with tunable reasoning_effort (xhigh/medium/low) and preserve_thinking to retain reasoning history — useful for multi-turn agentic workflows.
• Integration and serving: released artifacts are compatible with Transformers, vLLM, SGLang, TokenSpeed and similar engines; recommended sampling presets and API examples are provided to simplify production deployments.
• Benchmarks (high-level): shows substantial gains over prior 27B releases on coding and agentic tasks (examples: IFBench ~79.5, LiveCodeBench/LiveCodeBench-like scores high, MathVision with CI ~94.6, OmniDocBench ~91.1), indicating strong instruction following, multimodal reasoning and code/agent performance.

Who it's for and tradeoffs

Great fit if you need a compact, multimodal model that: integrates vision+text/video; runs agent-style multi-step tasks with preserved internal reasoning; and must be deployed with modern inference engines to balance throughput and cost. It’s well suited to developers building long-horizon agents, multimodal research, and production services that require large context windows.

Look elsewhere if you need the absolute top-of-stack accuracy regardless of serving cost (larger Max models may still be stronger in some benchmarks), or if your serving environment cannot support FP8 artifacts or the recommended inference frameworks. Also expect higher infrastructure demands when using the largest context lengths (memory and engineering to enable YaRN/RoPE overrides).