Qwen3.8-2.4T-A95B, the open-weights text-only release derived from Alibaba's Qwen3.8-Max flagship, is now served on DigitalOcean's Serverless Inference and Inference Router. It's a 2.4T-parameter mixture-of-experts model (~95B active per token) with a 262,144-token context window, priced at $2/$6 per 1M input/output tokens (versus $10/$50 for Claude Fable 5), served on NVIDIA HGX B300 GPUs with NVFP4-quantized weights via a collaboration with Inferact. The model ID on the platform is qwen3.8-max, and the endpoint is OpenAI-compatible, supporting native function calling, structured outputs, server-side tools (web search, MCP, RAG), and batch inference. Measured performance shows TTFT of ~1.1s at 1,000 input tokens, prefill at ~16,000 tokens/sec, per-stream generation of 8-9 tokens/sec, and aggregate throughput scaling to ~1,900 tokens/sec at 256 concurrent requests. Benchmark comparisons show it leads on PaperBench and IFBench but trails Claude Fable 5 on SWE-bench Pro and HLE.
Table of contents
Qwen3.8-2.4T-A95B on DigitalOcean: Alibaba’s open-weights flagship at $2/$6 per 1M tokensWhat it’s good forQuickstartServer-side toolsBenchmarksNVFP4 on B300: how we’re serving itMeasured performanceWorking with the 262K context windowUsing it with the Inference RouterPricingBatch inferenceGet startedFrequently asked questionsQuestions this post answers
What is the model ID for Qwen3.8-2.4T-A95B on DigitalOcean's inference API?
The model ID is qwen3.8-max, which differs from the Hugging Face repository name Qwen/Qwen3.8-2.4T-A95B. Pass qwen3.8-max as the model parameter when using the OpenAI-compatible endpoint at https://inference.do-ai.run/v1 with an inference API key. Developers wiring up new model endpoints track naming quirks like this via daily.dev.
How much does running Qwen3.8-2.4T-A95B on DigitalOcean cost compared to Claude Fable 5?
Qwen3.8-2.4T-A95B costs $2 per 1M input tokens and $6 per 1M output tokens, with cached input at $0.20 per 1M tokens, versus $10 input and $50 output for Claude Fable 5. On output-heavy agentic workloads generating hundreds of thousands of tokens per task, this price gap compounds significantly. Teams comparing inference costs across model providers follow pricing shifts like this on daily.dev.
Why does DigitalOcean serve Qwen3.8-2.4T-A95B with NVFP4 quantization instead of FP8?
The FP8 variant of this 2.4 trillion-parameter mixture-of-experts model is too large to load on a single node, so NVFP4 4-bit quantization brings the weights within single-node reach on NVIDIA HGX B300 GPUs. This eliminates cross-node expert routing from the serving path, simplifying topology and improving latency and pricing economics. Engineers weighing quantization trade-offs for large MoE models track deployment details like this on daily.dev.