Reports: GLM-5.3 post-training lifts Terminal-Bench from 4.6 to 28.3

• GLM-5.3 kept GLM-5.2's base model, while WesRoth's report puts Terminal-Bench 3.0 at 4.6 to 28.3 and DeepSWE at 46.2 to 66.9.
• More compute was spent on long-horizon task environments, and a same-prompt landing-page run from the landing-page comparison claimed lower token use and cost for 5.3.
• The training stack pushed train-rollout logprob mean absolute error to 1.89×10⁻⁷ at 4K tokens, according to ZhihuFrontier's deep dive.
• Z.ai reports CyberGym rising from 77.2% to 84.5%, while rauchg's security evaluation argued the lower-cost model could make repository-wide defensive scans cheaper to run.

Z.ai’s research report describes an ML-infrastructure task that gives an agent clusters, storage, internal documentation, codebases, and experiment results, then requires a measurable speedup. Its Terminal-Bench protocol allows 600 agent turns and a 10-hour cap inside isolated task containers. The slime documentation describes the core split as Megatron training connected to SGLang rollout, with custom data generation on top.

Z.ai says every GLM-5.3 gain came from post-training: more environments, more task diversity, and more RL compute on the GLM-5.2 base. A single public test from the landing-page comparison also claimed a better result with fewer tokens.

Z.ai’s benchmark table reports these changes against GLM-5.2:

• Terminal-Bench 3.0: 4.6 to 28.3, a 23.7-point gain.
• DeepSWE v1.1: 46.2 to 66.9, a 20.7-point gain.
• AutomationBench v1.0.6: 26.2 to 48.2, a 22.0-point gain.

The report’s evaluation notes specify max reasoning effort, 400K context, and 128K maximum output for Terminal-Bench 3.0. Results are avg@3 across three rollouts, each in an isolated container built from the official task image and scored by its separate verifier. Tool Search was disabled; each rollout had a 600-turn, 10-hour limit.

WesRoth's report characterized the shift as post-training in longer, more realistic engineering environments. Z.ai’s task-pipeline description fills in the machinery:

1. Research agents turn task patterns from real work into runnable environments with hidden state and multi-step dependencies.
2. A judge agent attempts each task to establish solvability.
3. Verifiers are synthesized without the reference solution; solver trajectories identify and close reward shortcuts.
4. Only verifiers that pass oracle, no-op, and unsolved-state checks generate the binary reward used directly for RL.

Z.ai says the environment-generation and verification pipeline still needs meaningful human-in-the-loop work.

The slime docs describe a training side based on Megatron and a rollout side based on SGLang. Z.ai says GLM-5.3 added full numerical alignment between them and reduced average logprob discrepancy to the 1e-7 range.

The technical account attributes the result to matching details that normally vary between implementations:

• Preserve the FP32 residual sum through RMSNorm, exposing a single BF16 rounding boundary.
• Use SGLang forward kernels for operations including DeepGEMM, RMSNorm, SwiGLU, RoPE, sparse MLA, and router GEMM, while retaining stable training-side backward implementations.
• Match the full MoE path, including top-k ordering, probability application, and accumulation order; it also reports that unmatched query-head padding could move the gap from about 2e-7 on H100 to 0.027 on B300.

Z.ai’s security results show the biggest relative gains farther along the exploitation chain:

• CyberGym: 77.2% to 84.5%, a 7.3-point gain.
• ExploitBench: 24.4% to 54.4%, a 30.0-point gain.
• ExploitGym: 29 to 105 tasks in the two-hour budget, and 39 to 130 in the six-hour budget.

For CyberGym, Z.ai used max reasoning effort, no web tools, 128K maximum output, and single-run pass@1 over 1,507 containerized tasks. Its own comparison still places Fable 5 ahead on ExploitBench and both ExploitGym budgets.

At announcement, access was through the GLM Coding Plan and ZCode; Nathan Lambert’s launch analysis reported API availability was still forthcoming and public weights were due two weeks later. Z.ai said the delay was for safety evaluation and hardening, while the AI Gateway announcement listed GLM-5.3 as coming soon.

The API migration note removes thinking.type: "disabled". GLM-5.3 accepts reasoning_effort values of low, high, and max; clients switching from disabled thinking must first set thinking to enabled with low effort, or the request fails.