Zhipu GLM-5.5 Post-Training Aims for 100% RSI and 100T Tokens

GLM-5.5 Post-Training: Multiple Internal Models Merged into Next-Gen Base

Leaked intern profile data shows Zhipu is running end-to-end programming post-training for GLM-5.5 on three internal model sizes: 32B, 122B, and 256B. After post-training, these models are being merged into the next-generation base model. The work involves a quantitative finance training environment , code sandbox , and financial programming RSI (Recursive Self-Improvement) .

glm-5.5-flash Appears in OpenCode Registry

A model entry named glm-5.5-flash has surfaced on OpenCode's data page, indicating a lighter variant is also in development.

Context: Previous 3.5 Trillion Parameter Rumor

Earlier reports suggested Zhipu's next flagship could reach 3.5 trillion parameters , targeting a system codenamed "Fable". The current post-training records add concrete detail to that roadmap.

Post-Training Emphasis Consistent with GLM-5.3 Strategy

Tang Jie (唐杰) previously explained that GLM-5.3's upgrades came from expanding task environments and reinforcement learning while keeping the base model and architecture fixed. The new records show this philosophy continuing: coding sandbox and quantitative finance environments ground training in specific task scenarios.

Zhipu's Pivot to Programming After DeepSeek R1 Pressure

In May 2025, facing market pressure from DeepSeek R1, Zhipu adjusted its model roadmap to fuse reasoning, programming, and agent capabilities. Since then, programming has become a primary investment focus, with more resources allocated to reinforcement learning environments and data engineering for coding .

100% RSI Goal: From Coding Agent to Medical Agent

Another leaked document sets a research target of "100% RSI" . The plan: start with a Coding Agent closed loop, then extend to a Medical Agent , while exploring how programming agents can generate their own training data. This builds on earlier self-improvement work:

HarnessPro : Agents review execution logs, analyze failures and waste, then automatically modify their runtime framework. This improved task performance and reduced per-task cost by 26.1% .

GLM-5.3-driven Infra Agent : The model participated in designing, debugging, and optimizing the inference infrastructure that runs itself (GLM-5.3-Flash).

Scale Correctness: Predicting Large-Scale Training from Small-Scale Experiments

On October 31, Tang Jie introduced "scale correctness" : systematically verifying whether conclusions from small-scale experiments hold when scaled up. By comparing curves across different scales, the team aims to improve predictions of large-scale training performance.

Self-Purification: Internal Judgment for Autonomous Iteration

For RSI to work, the system must identify whether a modification yields genuine improvement. Tang Jie calls this capability "self-purification" — the ability to autonomously judge right from wrong, decide when to stop, and correct errors. This aligns with the earlier GLM-6.0 fully self-training concept where autonomous judgment is the core challenge.

Data Engineering: Past Scale and 100T Token Target

Tang Jie disclosed historical GLM data coverage:

Web pages: 1 trillion

Documents: 2 billion

Books: 50 million

Code repositories: 500 million

Video/audio items: 10 billion

Audio hours: 300 million

Each source undergoes specific processing: quality filtering, layout restructuring, code execution testing, speech transcription. Tang Jie describes the data recipe as "life" for pre-training personnel.

The next target: 100T tokens , roughly 3.5× the 28.5T used for GLM-5. However, high-quality public text supply is estimated below 30T tokens, making effective data expansion a critical challenge.

Inference Deployment Trade-off: Speed vs. Concurrency

The team considered a ~1000 tokens/second ultra-fast version, but under the described configuration a single machine could serve only one user, making cost prohibitive. They opted to lower speed to achieve higher concurrency.

Multimodal Exploration: GLM-Omni and Knowledge Transfer

A paper on GLM-Omni investigates why knowledge learned through image generation is hard to access via text QA. The team experiments with interleaved image-text training and a visual chain-of-thought (generate image first, then answer) to improve cross-modal knowledge transfer. The relationship between GLM-Omni and GLM-5.5 versions remains undisclosed.

Outlook: GLM-5.5 as Integration Point

The leaked three-tier post-training and base-model merge records make the next-generation model's development contour more concrete. From increased post-training investment to programming environments, internal model merging, and self-evolution research, the key question for GLM-5.5's release is whether these accumulations translate into real gains on complex programming and long-horizon tasks.