REDWOOD CITY, Calif., Jan. 31, 2026 /PRNewswire/ -- Zilliz, the company behind the leading open-source vector database Milvus, today announced the open-source release of its Bilingual Semantic Highlighting Model, an industry-first AI model designed to dramatically reduce token usage and improve answer quality in production RAG-powered AI applications.
This highlighting model introduces sentence-level relevance filtering, enabling AI developers to remove low-signal context before sending prompts to large language models. This approach directly addresses rising inference costs and accuracy issues caused by oversized context windows in enterprise RAG and RAG-powered AI deployments.
"As RAG systems move into production, teams are running into very real cost and quality limits," said James Luan, VP of Engineering at Zilliz. "This model gives developers a practical way to reduce prompt size and improve answer accuracy without reworking their existing pipelines."
Key Innovations and Technical Breakthroughs
- Bilingual relevance by design: Optimized for both English and Chinese, the model addresses cross-lingual relevance challenges common in global RAG deployments. It is built on the MiniCPM-2B architecture, enabling low-latency, production-ready performance.
- Sentence-level context filtering: Rather than scoring entire document chunks, the model evaluates relevance at the sentence level and retains only content that directly supports a user query before sending it to the LLM.
- Lower token usage, higher answer quality: Zilliz reports that sentence-level filtering significantly compresses prompt size while improving downstream response quality, helping teams reduce inference costs and improve generation speed in production environments.
Availability
The Bilingual Semantic Highlighting Model is available today as an open-source release. To learn more about the training methodology and performance benchmarks, visit the Zilliz Technical Blog.
Download: : zilliz/semantic-highlight-bilingual-v1
About Zilliz
Zilliz is the company behind Milvus, the world's most widely adopted open-source vector database. Zilliz Cloud brings that performance to production with a fully managed, cloud-native platform built for scalable, low-latency vector search and hybrid retrieval. It supports billion-scale workloads with sub-10ms latency, auto-scaling, and optimized indexes for GenAI use cases like semantic search and RAG.
Zilliz is built to make AI not just possible—but practical. With a focus on performance and cost-efficiency, it helps engineering teams move from prototype to production without overprovisioning or complex infrastructure. Over 10,000 organizations worldwide rely on Zilliz to build intelligent applications at scale.
Headquartered in Redwood Shores, California, Zilliz is backed by leading investors, including Aramco's Prosperity 7 Ventures, Temasek's Pavilion Capital, Hillhouse Capital, 5Y Capital, Yunqi Partners, Trustbridge Partners, and others. Learn more at Zilliz.com.
** The press release content is from PR Newswire. Bastille Post is not involved in its creation. **
Zilliz Open Sources Industry-First Bilingual "Semantic Highlighting" Model to Slash RAG Token Costs and Boost Accuracy
A state-of-the-art general-purpose model in its class, built through automated AI research and designed to advance AI and scientific discovery.
BEIJING, Oct. 29, 2026 /PRNewswire/ -- Apex Intelligence has released Apex-35B-V1.0, its first model. Built on Qwen3.6-35B-A3B and further developed through mid-training and post-training, the 35B general-purpose model incorporates extensive proprietary datasets and research trajectories.
According to Apex Intelligence's evaluations, Apex-35B-V1.0 leads the similarly sized models tested in research and development capabilities while maintaining strong general-purpose performance. On selected benchmarks, it is competitive with substantially larger models.
AI participated autonomously in the model's development, including data construction, experiment design, code execution, analysis of training results and training optimization. Apex-35B-V1.0 can perform coding, agent workflows, question answering and deep search. It can also carry out end-to-end computational research workflows in AI research, computational science and robotics simulation.
Benchmark performance
Across 21 benchmarks, Apex-35B-V1.0 outperformed the similarly sized Qwen3.6-35B-A3B on 20 and Recursive Flint (40B) on 18, tying Recursive Flint on one. Beyond its strong performance on research tasks, the model demonstrated broad competitiveness in agentic tasks, coding, and reasoning, and surpassed larger models on selected benchmarks. Further details are available in the technical article.
Tech Blog：https://www.apexin.com/apex-35b/
Building Apex-35B-V1.0 Through Recursive Self-Improvement
Apex Intelligence developed Apex Research, an automated research system designed to enable recursive self-improvement. The system identifies model weaknesses, generates hypotheses, runs experiments and turns findings into measurable improvements. It contributed to the development of Apex-35B-V1.0 through self-evolving training data, automated experimentation, training recipes, GPU kernels and inference optimization.
Apex Research generates targeted training data to address weaknesses revealed by evaluation. When evaluations identified failures in computational fluid dynamics, the system analyzed the reasoning gaps, synthesized and validated new scientific training examples, and incorporated them into subsequent training rounds. Apex Intelligence reports that this process improved performance on its internal CFD-Bench by 32.1 points.
The system also analyzes evaluation results and agent trajectories to identify recurring failure patterns and refine training strategies. After identifying excessively long reasoning contexts and imbalanced data mixtures as factors affecting agent performance, it explored sequence-length distributions, short-to-long trajectory sampling ratios and token-aware batch compositions to improve agent reliability and training efficiency.
Apex Research extends this optimization to training and inference infrastructure. On the training side, its work on GPU load balancing, communication and computation overlap, dynamic batching, and token-aware workload distribution delivered 45–60% higher training throughput, according to Apex Intelligence. On the inference side, optimizations to attention kernels, KV-cache management and prefill/decode execution delivered 1.6–1.8 times faster decoding and 30–35% lower time to first token under comparable workloads.
Together, these improvements across data, training and infrastructure form a feedback loop in which each research iteration supports the next.
Availability
The weights for Apex-35B-V1.0 are available on Hugging Face. Accompanying code, a deployment guide and usage examples are available on GitHub. Developers can download and run the model for AI development, scientific computing and other research tasks. Model versions, hardware requirements, supported inference frameworks and licensing information are detailed on the project pages.
GitHub：https://github.com/apexin-ai/Apex-35B-V1.0
Hugging Face：https://huggingface.co/Apexintelligence-AI/Apex-35B-V1.0
About Apex Intelligence and what comes next
Apex Intelligence was founded by Yongchao Chen, an assistant professor at Tsinghua University's College of AI. The company aims to build the next generation of self-evolving foundation models for research. By "self-evolving," it means enabling AI to participate in research, experimentation and validation, then using the resulting discoveries to improve the model itself.
Apex Intelligence aims to advance AI from an assistant to a research collaborator capable of contributing to new discoveries. Its long-term vision is to develop models whose collective research capability matches and progressively surpasses that of 100,000 leading AI scientists across disciplines.
Following Apex-35B-V1.0, the company plans to release a larger self-evolving foundation model with stronger AI research capabilities and expanded applications in scientific fields including chip design and quantum computing. Apex Intelligence plans to announce further details on the model's capabilities, benchmark results and availability.
** This press release is distributed by PR Newswire through automated distribution system, for which the client assumes full responsibility. **
Apex Intelligence Releases Apex-35B-V1.0, a General-Purpose Model Built With Automated AI Research
