GLM-5.3: the open-weights coding flagship bought with post-training alone

In one line: the open-weights coding flagship bought entirely with post-training

GLM-5.3 is a 744B-A40B MoE that Z.ai (Zhipu) obtained by post-training the same base model as GLM-5.2, with weights published in FP8 and BF16 on HuggingFace and ModelScope. The companion release GLM-5.3-Flash is a 320B-A18B model on a newly trained base. The official claim is a single sentence: the most capable open-weights model for coding, 50 percent better than 5.2 on the in-house Z.ai Code Bench, and open-source SOTA on Terminal Bench 3.0 and Agents' Last Exam.

Same base, post-training only is the most information-dense fact in this release: not one byte of pre-training changed, every gain came from the post-training recipe and compute. It turns how much headroom post-training still holds into a measurable question, and it shows every open team chasing the closed frontier a route that does not require re-training a base model.

The data surface: two weight tiers, two bases, one deployment matrix

Weights	Size	Precision	Source and notes
GLM-5.3	744B-A40B	FP8	HuggingFace zai-org/GLM-5.3, ModelScope ZhipuAI/GLM-5.3
GLM-5.3-BF16	744B-A40B	BF16	Same, for teams that quantize themselves
GLM-5.3-Flash	320B-A18B	FP8	New base, sparse plus linear hybrid attention, mHC
GLM-5.3-Flash-BF16	320B-A18B	BF16	Same

The deployment matrix covers SGLang, vLLM, Transformers (glm_moe_dsa and glm5_next), KTransformers and Unsloth; Flash additionally lists TokenSpeed plus the Ascend side with vLLM-Ascend and xLLM. Fine-tuning goes through slime v0.3.0+ or ms-swift v4.4.0+. The FP8 weights of a 744B model do not fit one machine, but 40B active parameters keep the throughput arithmetic payable, which is exactly the intent DSA sparse attention has carried all along: bend long-context cost into a shape that can actually be served.

Reading the six benchmarks: two were claimed, four were beaten

Benchmark	GLM-5.3	GLM-5.2	Kimi K3	Fable 5	GPT-5.6 Sol
Terminal Bench 3.0	28.3	4.6	17.4	33.7	34.6
DeepSWE	66.9	46.2	67.5	69.7	72.7
Agents' Last Exam (CLI)	28.5	23.8	27.6	23.8	28.6
AutomationBench	48.2	26.2	46.7	46.2	45.8
HLE w/ Tools	62.5	54.7	59.8	63.9	64.5
GDPVal-AA v2	1769	1508	1682	1743	1730

Four readings. First, the open-source SOTA claim in the README circles only Terminal Bench 3.0 and Agents' Last Exam, and on those two the chart does lead (28.3 against 17.4 for Kimi K3, 28.5 against 27.6), so claim and chart agree. Second, on DeepSWE the 67.5 of Kimi K3 edges out 66.9, so open-source first must not be extrapolated to all six rows; this site writes what the chart shows and stops there. Third, AutomationBench 48.2 and GDPVal-AA v2 1769 beat every closed-source comparison on the chart (Fable 5 and GPT-5.6 Sol), which is beyond the official claim and worth recording on its own. Fourth, the generational gap over GLM-5.2 is an order of magnitude: Terminal Bench 3.0 from 4.6 to 28.3, AutomationBench from 26.2 to 48.2. That slope, from post-training alone on an unchanged base, is the real news.

Flash: the tier that attacks the supply-side ruler

Benchmark	GLM-5.3-Flash	GLM-5.2	DeepSeek-V4-Vision-Exp	Claude Opus 4.8	GPT-5.6 Terra	Gemini 3.7 Flash
Terminal-Bench 2.1	84.3	81.0	83.9	85.0	87.4	85.8
DeepSWE v1.1	63.4	46.2	59.3	58.0	69.6	65.3
Agents' Last Exam	26.3	20.4	27.3	27.0	28.0	-
AutomationBench v1.0.6	48.8	26.2	38.8	41.0	37.2	52.3
HLE w/ Tools	55.3	54.7	55.1	57.9	-	-
GDPVal-AA v2	1773	1504	1675	1582	1571	1527

The point of Flash is not topping the chart but the architecture: for the first time in the GLM series, sparse attention and linear attention live in one hybrid architecture, sharply cutting long-context serving cost while keeping precise long-context capability; mHC (Manifold-Constrained Hyper-Connections) lifts scaling efficiency on top of a 30T-token multimodal pre-training corpus. On the chart it takes Terminal-Bench 2.1 84.3, AutomationBench 48.8 and GDPVal-AA v2 1773, leading GLM-5.2 across the board and trading wins with the closed frontier. It answers a different question: for the same intelligence, how much compute and VRAM per token.

Lineage: three base changes, two ruler changes

Version	Size	Pre-training	Signature step
GLM-4.5	355B-A32B	23T	The MoE foundation
GLM-5	744B-A40B	28.5T	DSA plus slime async RL; open-source first on Vending Bench 2 (4,432 dollars)
GLM-5.1	Same base as 5	-	SWE-Bench Pro SOTA; leads NL2Repo and Terminal-Bench 2.0
GLM-5.2	Same base as 5	-	Solid 1M context (IndexShare); TB2.1 81.0, SWE-bench Pro 62.1
GLM-5.3	744B-A40B	Same base as 5.2	Pure post-training; Terminal Bench 3.0 28.3, open-weights first
GLM-5.3-Flash	320B-A18B	30T multimodal	Sparse plus linear hybrid attention, mHC, attacks serving cost

The cadence reads cleanly: 4.5 lays the foundation, 5 changes the base and adds DSA with async RL, 5.1 takes SWE-Bench Pro SOTA, 5.2 makes 1M context solid (IndexShare reuses the indexer every 4 layers, cutting per-token FLOPs by 2.9x at 1M and lifting MTP speculative accept length by 20 percent), 5.3 closes the coding and long-horizon gap on the same base with post-training alone, and Flash changes the base once more to attack supply. Technical report at arXiv 2602.15763; IndexShare at arXiv 2603.12201.

Deploying and reproducing: three traps that bite

First, reasoning_effort accepts low, high and max, and defaults to max when omitted (any other value also falls back to max); reproducing the boards requires keeping the default max, while GLM-5.2 accepts only high and max, so cross-version scripts need a branch. Second, in the chat template clear_thinking defaults to false when omitted, so a multi-turn conversation that does not pass true explicitly carries the previous turn thinking chain into context, changing both the token bill and the behavior. Third, weights and inference stack must be paired: 5.3 runs on glm_moe_dsa, 5.3-Flash on glm5_next, and running new weights under an old configuration tends to fail silently rather than raise.

Emergent cyber capability: why this is also evidence for the safety domain

The vendor itself admits cyber capability developed faster than expected: as post-training scaled, GLM-5.3 reached SOTA on CyberGym for vulnerability discovery, with the largest gains further up the exploitation chain, more than doubling GLM-5.2 on exploitation benchmarks, on capability that was never trained separately. Emergent capability cannot be predicted from a training manifest, so it cannot be recalled by one either; the attack surface of open weights grows with post-training scale, and evaluation plus disclosure cadence has to keep up. This release therefore also feeds the evidence stream of the safety domain.