GLM-5.3-Flash outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price, while approaching Claude Opus 4.8 on coding and agentic benchmarks.
For the first time, a hybrid architecture combines sparse and linear attention, sharply reducing long-context serving costs. Manifold-Constrained Hyper-Connections (mHC) improve scaling efficiency. A 30T-token multimodal pre-training corpus lets it produce more intelligence with less compute.
Before release, we tested it anonymously as ox-alpha on OpenCode and OpenRouter. It quickly became the most popular model of the week — all traffic served on Chinese AI chips.
Competitive performance,
flash cost
GLM-5.3-Flash pushes the Pareto frontier of the Artificial Analysis Intelligence Index v4.1.1, scoring 57 at $0.045 per task — intelligence previously available only at roughly 10× the cost.
Across six coding and agentic benchmarks it consistently outperforms GLM-5.2: 63.4 vs 46.2 on DeepSWE v1.1, 48.8 vs 26.2 on AutomationBench — while approaching Claude Opus 4.8 overall.
On Z.ai Code Bench at max effort: 29.0 vs 29.5 — a near tie with Opus 4.8.
Architecture for
extreme efficiency
Despite a similar total parameter count (320B vs 355B), GLM-5.3-Flash nearly halves the activated parameters (18B vs 32B) and layers (45 vs 92) versus GLM-4.5.
Linear attention captures local dependencies through state modeling; sparse attention retrieves global context through a lightweight indexer. IndexPool compresses four indexer key vectors into one via weighted pooling.
Attention compute 3.0× smaller. KV cache 4.4× smaller. Lowest attention compute of all models compared.
| Benchmark | GLM-4.5-Base | GLM-5-Base | DS-V4-Flash-Base | GLM-5.3-Flash-Base |
|---|---|---|---|---|
| Specs | ||||
| Activated Params | 32B | 40B | 13B | 18B |
| Total Params | 355B | 744B | 284B | 320B |
| Base model results | ||||
| MMLU | 86.1 | 88.3 | 88.5 | 88.1 |
| BBH | 86.2 | 87.4 | 84.9 | 86.6 |
| HellaSwag | 87.1 | 88.1 | 85.3 | 87.1 |
| LiveCodeBench-Base | 28.1 | 34.4 | 29.9 | 37.6 |
| SimpleQA | 30.0 | 36.0 | 31.2 | 33.5 |
Code lets the model build and change the world. Vision lets it enter the world people see and use.
Visual intelligence
in the coding loop
Visual coding expands what code can reach. For frontend, game development, and 3D simulation, the output is not code alone — it's an interface, an interaction, a world. Many failures only surface through rendering, interaction, playtesting.
Vision is natively integrated: the model decides when to observe and uses visual feedback to guide actions. Its pipelines demand self-visual judgment — interacting with environments, inspecting outputs, refining iteratively.
Validation beyond functional correctness — the rendered, interactive product.
Beyond coding —
your partner at work
Professional work is interpreting heterogeneous visual and structured information — documents, spreadsheets, presentations, dashboards. Rather than translating the environment into text, the model directly interprets the artifacts of the task, assesses its own output against visual context, and refines.
Stronger judgments of presentation quality and aesthetics — self-verification that sees what the user sees.
Serving at scale on
Chinese AI chips
A dedicated inference engine on top of SGLang — accelerated by the GLM-5.3-powered infrastructure agent. A feedback loop where the model helped optimize the system serving the model itself.
Intra-node tensor parallelism, ReplaySSM, W8A8 quantization, hybrid INT8/FP8/BF16 cache quantization, Layer Split — plus Encode–Prefill–Decode disaggregated architecture at cluster scale.
Full benchmark
results
| Benchmark | GLM-5.3-Flash | GLM-5.2 | DS-V4-Vision | Opus 4.8 | GPT-5.6 Terra | Gemini 3.7 Flash |
|---|---|---|---|---|---|---|
| Coding | ||||||
| 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 |
| NL2Repo | 56.3 | 48.9 | 57.7 | 69.7 | — | — |
| Agentic | ||||||
| Toolathlon Verified | 78.4 | 59.9 | 75.9 | 76.2 | 74.9 | — |
| AutomationBench v1.0.6 | 48.8 | 26.2 | 38.8 | 41.0 | 37.2 | 52.3 |
| Agents' Last Exam | 26.3 | 20.4 | 27.3 | 27.0 | 28.0 | — |
| HLE w/ Tools | 55.3 | 54.7 | 55.1 | 57.9 | — | — |
| GDPval-AA v2 | 1773 | 1504 | 1675 | 1582 | 1571 | 1527 |
| Vision | ||||||
| OfficeQA Pro | 62.4 | — | 57.9 | 48.9 | — | — |
| CharXiv Reasoning | 89.4 | — | 80.4 | 89.9 | 88.0 | 88.7 |
| Chartography w/ Tools | 78.0 | — | 64.3 | 75.0 | 68.0 | 65.0 |
| BabyVision | 53.4 | — | 35.1 | 46.8 | 61.6 | 70.9 |
| MVbench | 77.8 | — | 69.4 | 67.1 | 75.0 | 82.2 |
| MMVU | 80.5 | — | 72.7 | 67.4 | 75.8 | 82.3 |
Footnotes & evaluation protocols
HLE w/ tools: temperature=1.0, top_p=0.95, max 163,840 tokens, context 300K. Judge: GPT-5.6-luna (medium).
DeepSWE: mini-swe-agent harness, temperature=0.95, timeout 6h, 400K context.
Terminal-Bench 2.1: Claude Code 2.1.207, temperature=1.0, max 6h timeout.
Toolathlon: official evaluation service, pass@1 averaged over 3 runs.
BabyVision: shorter side ≥1.5K pixels.