Z.ai releases 743B-parameter GLM-5.3 open weights

• Z.ai has released GLM-5.3 weights for download and customization. The FP8 checkpoint is roughly 744B parameters with 40B active per token, according to vllm_project's launch note, and Zai_org's announcement calls it its most capable model for agentic coding and cyber defense.
• The model keeps GLM-5.2's base and derives its gains from post-training, while the official model card reports Terminal Bench 3.0 rising from 4.6% to 28.3% and ExploitBench from 24.4% to 54.4%.
• The weights arrived after a two-week safety evaluation: Zai_org's release-date post set the August 28 download date, while a DeepLearning.AI report ties the hold to its vulnerability-finding and exploitation capability.
• Serving support landed immediately across vLLM, SGLang, Baseten, OpenRouter, Tinker, Ollama, Together and Modal. lmsysorg's SGLang announcement covers NVIDIA and AMD hardware, while baseten's API notice offers a US-hosted, zero-data-retention endpoint.
• “Local” still means serious hardware: kimmonismus's hardware inventory estimates 10 to 12 H100s or eight H200s for FP8, with aggressive two-bit variants trading quality and context for a 230 to 250 GB footprint.

The official model card makes its Terminal Bench 3.0 setup unusually reconstructable: Claude Code 2.1.207, 400K context, 600 agent turns, three rollouts, and a ten-hour cap per rollout. vLLM retained its GLM-5.2 serving path, parsers and MTP checkpoint handling, according to vllm_project's deployment note. The GLM-5.3 license screenshot also defines a high-revenue model-hosting carveout, an odd wrinkle in an open-weights release.

• Weights: an FP8 GLM-5.3 checkpoint listed as 744B-A40B by vllm_project's launch note, alongside download-and-customize access from Zai_org's announcement.
• I/O: text input up to 1M tokens and text output up to 128K tokens, per the DeepLearning.AI report.
• Controls: reasoning_effort accepts low, high, or max, defaulting to max; the official model card says chat callers must explicitly set clear_thinking=true to clear reasoning.
• API pricing: $1.40 input, $0.26 cached input, and $4.40 output per million tokens in the DeepLearning.AI report.
• Post-training: Baseten announced day-zero Loops support in baseten's Loops announcement, and oneill_c's post framed the smaller-than-Kimi-K3 footprint as a cheaper base for RL, SFT, or OPD work.

First-party

• Terminal Bench 3.0: 4.6% → 28.3%, +23.7 points, per the official model card linked by Zai_org's release post.
• DeepSWE v1.1: 46.2% → 66.9%, +20.7 points, in the official model card behind Zai_org's announcement.
• CyberGym: 77.2% → 84.5%, +7.3 points, according to the official model card released in Zai_org's post.
• GDPval-AA v2: 1508 → 1769 Elo, +261 Elo, per the official model card attached to Zai_org's weights release.

Third-party evaluators

• Artificial Analysis Intelligence Index, max reasoning: 53 → 60, +7 points, according to the DeepLearning.AI assessment of the model in Zai_org's announcement.

Customer-reported

• Perplexity Computer's WANDR: 0.259 → 0.278, +0.019, in perplexity_ai's WANDR report.

The public-weight release followed the two-week review period. The DeepLearning.AI report says Z.ai evaluated the model with vetted security partners after its exploit skills scaled faster than expected.

ExploitBench still puts GLM-5.3 at 54.4%, 23.6 points below Fable 5's 78.0% and 22.1 points below GPT-5.6 Sol's 76.5% in the official model card. The same DeepLearning.AI report lists weaker general-knowledge results, including 42.3% on Humanity's Last Exam and 34% AA-Omniscience Accuracy.

The flagship is text-in, text-out. Native multimodality belongs to the 320B-A18B GLM-5.3-Flash sibling, as Zai_org's Flash announcement specifies. emollick argued that releases at this capability level need model cards and red teaming because open weights remove hosted guardrails.

GLM-5.3 shares GLM-5.2's base model, according to the official model card. The release is a scaled post-training result rather than an architecture refresh.

The training mix expanded from coding puzzles toward simulated workspaces with codebases, documentation, clusters, storage and test tools. The DeepLearning.AI report says Z.ai used single-rollout asynchronous optimization with compaction, had agents generate some environments and reward signals, and checked grader agents against both completed and untouched tasks before accepting their scores.

The model card exposes the operational knobs and evaluation assumptions:

• max is the default reasoning setting, even for an invalid reasoning_effort value.
• Terminal Bench 3.0 uses 400K context and a 128K output cap.
• Agents' Last Exam uses 1M context and a 64K output cap.
• ExploitGym rescales API time by model throughput, using 115 tokens per second for GLM-5.3.

A comparison from rohanpaul_ai put the same three prompts through the API alone and Atomic Agent. The harness run consumed 1,091,380 tokens for $3.16, versus 543,290 for $2.39, while keeping state, executing actions and feeding failures back into the loop.

An Atomic Chat Blender task ran both siblings through roughly 40 minutes of MCP tool use. rohanpaul_ai's Blender test reported 811 objects for Flash at $0.0526 and 847 for full GLM-5.3 at $0.8807.

abacaj said in a hands-on note they use Flash for code, tool calling and browser driving, and rated its high-effort mode comparable to Opus 4.8. The counterpoint came from bridgemindai, who wrote in a criticism that local Flash models benchmark well but still lack polish and reliability in real production work.

Self-hosted runtimes

• vLLM serves GLM-5.3 with vllm serve zai-org/GLM-5.3 -tp 8; vllm_project says it reuses the GLM-5.2 path, glm47 and glm45 parsers, MTP checkpoint support, and FP8 KV cache through 1M context.
• SGLang retained its runtime and flags across NVIDIA Blackwell and Hopper plus AMD MI300X, MI325X and MI355X, according to lmsysorg's SGLang release. SGLang also powers Slime, the rollout engine Z.ai used for post-training.

Hosted endpoints

• Baseten exposed a US-only, zero-data-retention Model API in baseten's launch post and paired it with the Loops post-training stack in baseten's Loops announcement.
• OpenRouter listed the full model for complex software engineering, long-horizon agents and cybersecurity in OpenRouter's availability post. Tinker exposed a 256K-context endpoint in tinkerapi's deployment note.
• Modular Cloud, Together and Modal each announced availability in clattner_llvm's Modular Cloud post, togethercompute's launch, and modal's availability post.
• Ollama placed both GLM-5.3 and Flash in its US and Europe cloud in ollama's rollout, with commands for Claude Code and OpenCode harnesses. Perplexity Computer added GLM-5.3 for long-context multimodal agent workloads in perplexity_ai's availability report.

The release uses a GLM-5.3 license, not the MIT terms used by Flash. The text permits use, modification and redistribution, but requires a Z.ai security review before commercial use by a Model-as-a-Service operator whose own and affiliates' aggregate revenue exceeds $10 billion in any rolling 12-month period. Embedded end-user features and mere request relays are explicitly excluded from that definition.

The FP8 load remains a data-center deployment. kimmonismus's hardware inventory lists 10 to 12 H100s or eight H200s for FP8, roughly 390 to 430 GB for four-bit or NVFP4, and roughly 230 to 250 GB for aggressive two-bit quantization. UnslothAI's two-bit release reports a 239 GB GGUF at about 81% accuracy after shrinking the model from 1.51 TB, putting a concrete quality number on the smaller local footprint.