Alibaba's new open-source AI matches top paid AI models on agent work, but running it locally costs hours, not dollars

Alibaba's Qwen team released a 27-billion-parameter open-source AI on August 14 that, on local test runs, can now do coding and agent work in the same league as the top paid frontier models. The model, Qwen3.8-27B, ships under an Apache 2.0 license, a free, permissive license that lets anyone download, modify, and redeploy it without paying Alibaba, and can be loaded onto a single high-end server. The catch surfaced in a hands-on review by Lei Feng (雷锋网): the run that delivers that quality takes 6 hours and 17 minutes of wall-clock time and burns about 14 million tokens. The new bottleneck for open-source is not the API bill. It is the clock on the desk.

The Lei Feng team ran the model on a single A100 server, a high-end Nvidia GPU widely used for AI training and inference, across three tracks. First, standard deployment using three popular open-source serving frameworks (vLLM, SGLang, and Llama.cpp) measured stability and throughput on seven standard tasks covering reasoning, instruction following, chitchat, news writing, and result attribution. Second, quantization, a technique that shrinks the model's memory footprint by storing its weights at lower numerical precision (from 16 bits per weight down to 2 bits, trading some quality for size), was tested across the 2-to-16-bit range. Third, an agent scenario plugged the model into DeepSeek Harness, a workflow tool that lets a model autonomously break a goal into steps and run them in a loop, with a locally-deployed DeepSeek-V4-Flash-0731 as a control, across nine agent tasks including paper review and a 3D-website build.

vLLM was the most stable of the three frameworks, completing all 14 runs without directional errors. vLLM and SGLang both averaged above 40 tokens per second of generated text, roughly 60 words a second, fast enough for interactive chat. Quantization held up better than expected: 3-bit to 6-bit settings preserved full quality on the text task set, and 2-bit and 16-bit settings produced documented tradeoffs in file size, speed, and quality.

The agent track is where the cost shape changed. Across all nine agent tasks, Qwen3.8-27B scored full marks on task completion. It was not failing on quality. The run consumed 13,995,350 tokens, fired 197 model requests, and lasted 22,564 seconds (6 hours and 17 minutes). One task, generating a 3D website, accounted for 11,533,959 of those tokens by itself, more than 80% of the total. The reason, per the Lei Feng root-cause analysis, is not a model defect. It is workflow cost: oversized task decomposition (the model broke the goal into too many steps), single-step goals that carried too much context, repeated context reinjection on every iteration, and long idle time waiting for tools. The harness is the bottleneck, not the weights.

A single 3D-website build, run end-to-end, takes the better part of a working day on one A100. A nine-task agent suite is a workday and a half. Multiply that across a team and "free" starts to look like payroll. The open-source story has been promising cheaper. The actual bill is now operator time.

The honest counterargument is that this is a single test rig on a single seed. A multi-server, multi-seed reproduction could compress the wall-clock by an order of magnitude. The Lei Feng numbers are illustrative, not load-bearing for the model class. The shape of the cost, heavy on context, decomposition, and idle time on a single machine, is what a reader can plan around. The exact 6 hours and 17 minutes is what one operator got on one server.

The Lei Feng team concludes that the model needs strict tokens, steps, and output-boundary constraints to be useful in real agent pipelines. The operator's job is now to engineer the harness: bound the context window per step, cap the iteration count, batch tool calls, pre-stage assets so the model is not redrawing the same ground on every loop. The model is good. The workflow is the part that has to be written.

Alibaba's release dropped on August 14. The Lei Feng hands-on is the first public empirical measurement of what running that class of open-source model actually costs under a realistic agent workload. Open-source in 2026 is not cheaper in the way wire coverage keeps promising. It is free in license and expensive in clock, and the gap between the two is where the next round of tooling will sit.