Open-Source AI Agents Show Promise but Face Productization Hurdles, Tests Find

OpenWorker, released on July 23 by Andrew Ng and Rohit Prasad as a desktop-based open-source agent, drew developer scrutiny within its first week. Leiphone reports that users who tried to connect OpenAI-compatible endpoints or local Ollama services encountered missing-model errors even after passing configuration checks. Community feedback also raised concerns about data-flow transparency and approval mechanisms that may not cover all execution paths, with some configuration files reportedly starting external service processes before a workspace was explicitly trusted.

For Alibaba's Qwen3.8-27B, Leiphone ran standard, quantized and agent-benchmark tests. In standard deployment, vLLM and SGLang both delivered average throughput above 40 tokens per second, while Llama.cpp reached 23.5. Quantization from 3-bit to 6-bit retained perfect scores on seven text tasks, while 2-bit showed factual-correction errors. However, when connected to the DeepSeek Harness for nine agent tasks, the model scored full marks on task completion but consumed 13,995,350 tokens and took about 6 hours and 17 minutes; a single 3D website task accounted for 11.5 million tokens.

A separate evaluation compared Qwen 3.8-Max and GPT-5.6 iterating an existing 3D "AI universe" web project. Qwen 3.8-Max, running through Qwen Code, spent 3 hours 29 minutes and sent 195 model requests, producing a visually richer result but exhausting its token plan before completing acceptance checks. GPT-5.6, via Codex, finished in 37 minutes with a working but less polished output. Leiphone noted the results represent each model's performance in its official agent environment.

Zhipu's GLM 5.3 and GLM 5.2 were both tasked with building a single-player 3D open-world driving game for Beijing's Guomao-Wangjing area using OpenStreetMap data. Both models delivered runnable projects, with GLM 5.3 finishing in 50 minutes versus 66 minutes and adding more realistic lighting, streetlights and navigation. However, Leiphone found that GLM 5.3's strengthened safety mechanisms frequently triggered refusals during automated toolchain interactions, occasionally interrupting normal development flows.

DeepSeek's open-weight release of V4-Flash-Vision-Exp revealed a vision pipeline that goes beyond adding an image encoder. Images are processed by a 32-layer ViT, compressed by a 3x3 aligner, and inserted into the V4 token sequence with special attention and MoE routing rules. According to Leiphone's code analysis, this design keeps visual tokens distinct from text tokens and allows the model to handle spatial completeness and expert selection. But in agent loops that repeatedly observe an environment, near-identical screenshots are re-encoded every time, potentially creating a computational bottleneck.

These developments come as Chinese AI application startups face increasing pressure from both big tech and model upgrades. Leiphone described investor Lin Wang's metaphor of "building houses on quicksand," where a single model update can erase an application's hard-won value. Founders like Chen Mian (Lovart), Jing Kun (Genspark) and Ni Zhengmin (Flowith) have chosen distinct routes, from embedding AI into professional workflows to expanding into broader task entry points or building long-term agent systems.