Two open models are taking different routes to useful AI. Fastino Labs has released GLiNER2.5-Decide, a 340M-parameter open-weight decision model, while China Telecom Artificial Intelligence Technology Co., Ltd. has released Xing4.0-29B-A4B, a lightweight model built for agentic work.
They are not competing for the same job. GLiNER2.5-Decide handles structured operational decisions without generating prose, while Xing4.0-29B-A4B targets multi-step tasks that involve planning, tools, code, and long documents.
Fastino Builds a Decision Specialist
GLiNER2.5-Decide takes text and a schema of typed questions, then returns structured answers with a probability distribution, confidence score, and constraint-feasibility metadata. That design makes it a classifier for operational decisions, not another chatbot waiting to explain everything at length.
The model uses a DeBERTa-v3-large encoder and was fine-tuned from gliner2-large-v1. Fastino is clear about the boundaries: GLiNER2.5-Decide does not reason, explain, or answer open questions. It makes constrained decisions, which is a narrower job and, in many systems, a more useful one.
Fastino evaluated the model on Fast Decisions, a test suite containing 5,100 examples across 17 datasets. GLiNER2.5-Decide led 9 of the 17 datasets, giving the compact model a result that deserves attention beyond its parameter count.
The latency figures make the deployment case stronger. At batch 1, using a 2-head, 15-label schema and 64 tokens, end-to-end latency reached 167.3 ms on a 48-vCPU Intel Xeon Platinum 8581C, 43.6 ms on an NVIDIA T4, 43.4 ms on an NVIDIA L4, 38.3 ms on an NVIDIA V100, and 47.3 ms on an NVIDIA A100.
Fastino also measured the model at 1,024 tokens, where latency was 52.6 ms on an NVIDIA A100, 75.6 ms on an NVIDIA V100, and 131.4 ms on an NVIDIA L4. Those numbers position GLiNER2.5-Decide as a candidate for CPU and modest-GPU workflows rather than a model that demands a server room and a ceremonial budget meeting.
The release includes two related models. GLiNER2.5-Decide-1B scores 59.6% on the same suite, while GLiNER2.5-multi-Decide is a 287M model for multilingual input that scores 56.7%.
Xing4.0-29B-A4B Takes the Agent Route
Xing4.0-29B-A4B takes the opposite approach. Its Mixture-of-Experts architecture contains 29 billion total parameters, with 4 billion activated parameters, and it can plan multi-step task paths, call external tools, process long-context inputs, and deliver completed results.
Its context window reaches 256K tokens, giving the model room to work across long documents and large software projects. China Telecom AI says the model can break down high-level goals, analyze multi-file repositories, execute code, and generate structured deliverables for complex engineering scenarios.
The hardware requirement is the more interesting part. Xing4.0-29B-A4B requires 15 GB of GPU memory with low-bit quantization and memory optimization techniques, and it can run locally on a single consumer-grade graphics card. A 29-billion-parameter model that fits this profile is a different proposition from an agent that requires access to an expensive cluster.
On SWE-bench Verified, Xing4.0-29B-A4B scored 75.0 out of 100. The model has also been deployed by China Telecom in enterprise environments, including customer service platforms and interactive service scenarios, giving its agentic claims a production setting rather than leaving them trapped in a demo.
China Telecom AI released Xing4.0-29B-A4B on GitHub under XingChen-AGI and on Hugging Face. The company plans to release larger-parameter models in the coming months, because apparently 29 billion parameters now counts as the warm-up act.
A spokesperson for China Telecom AI said, “We believe the future of AI lies not in ever-larger models, but in making powerful intelligence accessible to everyone, affordable, and deployable anywhere.” The two releases point in that direction from different angles: Fastino trims decision-making into a fast structured component, while Xing4.0-29B-A4B packages planning and execution into a model that can fit on local hardware.
Based on
- Fastino Releases GLiNER2.5-Decide: A 340M Open-Weight Decision Model That Runs on CPU — marktechpost.com
- Xing4.0-29B: A Versatile Agentic Large Model That Runs on a Single GPU, Now Officially Released | Markets Insider — markets.businessinsider.com
- China Telecom AI Officially Releases Xing4.0-29B Agentic Large Model for Single-GPU Deployment | Currency News | Financial and Business News | Markets Insider — markets.businessinsider.com
