Baidu releases Unlimited OCR 2 model for single-pass parsing of documents of any length

The Beijing Academy of Artificial Intelligence (BAAI) has unveiled FlagOS 2.2—the latest version of its open software stack designed to run artificial intelligence models across various chip architectures. This release was made on September 30th and demonstrated at the AI Open Computing Conference in Beijing on October 17, 2026. BAAI characterizes this update as a shift from adapting models to multiple chips to optimizing the general code for them.

The main achievement is the scale. FlagOS 2.2 covers nine operator libraries, and the number of public operators in FlagGems and related libraries has increased from 564 to 1454, an increase of 890 units. The new FlagGems-SGLang library added 40 multi-chip operators that cover functions such as attention, Mamba state updates, expert mixing, LoRA, sampling, and quantization. BAAI asserts that thanks to over a thousand operators across 10 types of AI hardware, overall operator performance has increased by approximately 27%.

A significant part of this growth is due to automation. BAAI's KernelGen operator generator now supports nine types of AI chips and created 765 new operators this quarter. Of these, 416 passed accuracy checks on at least five domestic chips, achieving over 80% of native operator performance. Furthermore, the unified debugging and profiling tool FlagPrism connects Triton source code data, compiler intermediate representation, and on-device execution data, supporting Huawei Ascend, Iluvatar CoreX, Moore Threads, Nvidia, and Enflame.

The release also includes FlagOS-Compressor for weight conversion and quantization between chips. BAAI reported that applying W8A8 to the Qwen3.5-397B-A17B model reduced the model file size from 751.41 GB to 382.45 GB, a decrease of 49.1%, while the model achieved 84.85% on GPQA-Diamond. Combined with framework plugins for vLLM, Megatron-LM, SGLang, and others, new models can be adapted for up to 12 chip platforms within just one day of release.

Research using older hardware demonstrates success. On Ascend 910B servers, BAAI noted that the logical serving throughput of the GLM-5.3-Flash model in agent reproduction tests increased 2.04 times compared to pre-optimization levels, rising from 8.52k to 17.36k tokens per second. The Xing4.0-29B-A4B model from China Telecom was adapted for 10 chip platforms, resulting in a 19.87% reduction in time-to-first-token in long-input tests compared to native operators.

BAAI believes the issue lies in timelines: over 80% of deployed AI computations relate to non-latest generation, while models are rapidly changing. It should be noted that all figures presented are based on BAAI's internal tests, and the company emphasizes that actual serving performance still depends on the workload.

At the 10th CSDI summit in Shenzhen on October 16, SenseCore Vice President and CTO of SenseTime, Xuan Shanming, discussed how the company is restructuring its inference infrastructure for the age of agents. In this new era, the demand for tokens significantly exceeds the capabilities of chatbots due to multi-agent workflows and long tasks performed without constant supervision.

The main focus of the presentation was not on the number of accelerators in the data center, but on the ability of this center to stably produce high-quality tokens within specified cost and service level constraints.

The SenseCore solution is based on separating the Prefill and Decode (PD) processes within heterogeneous clusters of 10,000 GPUs located in their AIDC in Lingang. The Prefill process handles input tokens and is computationally intensive, while Decode generates output tokens by constantly accessing the context cache and requiring high memory bandwidth.

In the SenseCore architecture, domestic accelerators are focused on executing Prefill, while high-performance GPUs handle Decode. These stages exchange state via KV Cache channels, which cover mixed factories with bandwidths of 800G, 400G, and 200G within a single cluster.

This separation creates an operational risk: in some configurations, three to fifty low or medium-level Prefill nodes may serve one Decode node, meaning that a failure of a Decode node could halt a large group of Prefill nodes. To minimize this risk, SenseCore maintains a small reserve of high-end cards in the Prefill pool for quick switching, and also allows nodes to dynamically change roles between Prefill and Decode depending on the load.

During one test scenario using the same number of flagship high-performance GPUs as the baseline configuration, the company reported approximately a fivefold increase in overall token throughput after implementing heterogeneous PD scheduling. Furthermore, the company is exploring cross-domain PD, directing Prefill work to regional medium-level capacities and Decode work to remote high-bandwidth clusters.

This becomes relevant against the backdrop of the decreasing size of the KV Cache for million-token contexts—it shrinks from nearly 100GB to about 1GB on some models, as well as due to the separation of Attention/FFN within the Decode process.

SenseCore claims that its platform currently processes about 6 trillion tokens daily and aims to reach 10 trillion tokens per day by the fourth quarter of 2026, with a projected annual growth in total tokens of approximately 25 times. It should be noted that these throughput and speed increases are company statements. During the presentation, it was emphasized that domestic chips are intended to take up a larger share in Prefill and mixed inference tasks, rather than completely replacing the high-performance GPUs used for Decode.