Z.ai releases 320B-parameter GLM-5.3-Flash, claiming near-Opus coding results at $0.045 per task
- Z.ai released GLM-5.3-Flash, its first natively multimodal GLM-5 model, with 320B total parameters and 18B active parameters; it reports an Artificial Analysis Intelligence Index score of 57 at a discounted cost of $0.045 per task.
- Across six coding and agent benchmarks, Z.ai reports GLM-5.3-Flash beat GLM-5.2, including 63.4 versus 46.2 on DeepSWE v1.1 and 48.8 versus 26.2 on AutomationBench; on its Code Bench at maximum effort, it scored 29.0 versus Claude Opus 4.8's 29.5.
- The model uses hybrid linear and sparse attention plus IndexPool, which pools four indexer key vectors into one; Z.ai reports 3.0x lower attention compute and a 4.4x smaller KV cache than GLM-5.3.
- Z.ai tested the model anonymously as ox-alpha on OpenCode and OpenRouter before release, says it became the week's most popular model, and says the traffic ran on Chinese AI chips.
- The base model scored 37.6 on LiveCodeBench-Base, above GLM-5-Base's 34.4 and DeepSeek-V4-Flash-Base's 29.9 in Z.ai's evaluation framework; its MMLU score was 88.1, versus 88.3 for GLM-5-Base.
Hacker News 의견들
At the listed API rates, input is $0.15 per million tokens, output is $0.50, and cached input is $0.03. Is that actually cheaper than DeepSeek V4 Flash?
I think US sanctions are pushing China to build more of its own technology. That outcome was predictable once export controls started pushing harder.
I am skeptical of the sanctions framing because China has a long record of technology theft. Still, that argument gets weaker when its work is published openly and can be replicated.
The Chinese-chip claim is the real story to me. Z.ai says it got a 3x end-to-end serving improvement on the same hardware and reached per-token costs comparable to mainstream NVIDIA GPUs.
I used Ox Alpha and it ran badly for me: about 20 tokens per second, very high latency, and frequent timeouts. The chips may be fine, but capacity clearly did not meet demand.
Ox Alpha was a smaller model and felt slow to me. Chinese accelerators are improving, but NVIDIA still has a very large lead.
I do not expect serious API customers to send sensitive data to a Chinese provider. The practical alternatives are local hosting or leasing compute, and both have cost or capacity problems.
I bought four Sparks and want to try this locally. My two-node DeepSeek V4 setup handles basic work, but it stalls on novel technical problems that need deeper reasoning.
I use local models and frontier models side by side, and the gap is still large on hard work. Claims that a local LLM replaces Opus-level models sound like wishful thinking, engagement bait, or tolerance for lower-quality output.
I initially got the memory math wrong: Q4 is about 160 GB, so 192 GB is a realistic minimum and 256 GB unified memory can run 5-bit quantization. It is still expensive home hardware.
I do not think GLM's use of "Flash" matches DeepSeek V4 Flash on speed. Throughput matters enough that the benchmark gains may not put them in the same segment.