GLM-5.3-Flash collected 641 Hacker News points in about two and a half hours, an unusually fast climb for a model release. The more useful number for developers sits outside that discussion: Artificial Analysis measured the model at only 48.7 output tokens per second. That makes Z.ai's new "Flash" model a curious proposition. It is cheap, open-weight and competitive in independent tests, yet speed and token consumption could erase part of the price advantage in agent workloads. The Hacker News thread had reached 299 comments by the time of writing.
Z.ai released GLM-5.3-Flash on August 26 as a 320-billion-parameter mixture-of-experts model with 18 billion active parameters. It accepts text and images, has a one-million-token context window and uses an MIT license. The Hugging Face model card lists local serving support through SGLang, vLLM, TokenSpeed and KTransformers. This is a downloadable model rather than an API wrapper whose weights remain hidden.
The low price has an expiry date
Z.ai currently charges $0.075 per million input tokens and $0.25 per million output tokens for GLM-5.3-Flash. Cached input costs $0.015 per million tokens, and cache storage is temporarily free. Those are promotional rates at half the listed prices. The company's pricing page says the offer ends at 24:00 on September 9, 2026, in Singapore time. After that, the posted list prices are $0.15 for input and $0.50 for output.
That is still far below Z.ai's own GLM-5.3 and GLM-5.2 APIs, each priced at $1.40 per million input tokens and $4.40 per million output tokens on the same price table. On the posted rates, one million generated tokens from GLM-5.3-Flash cost one-eighteenth as much as the same output volume from GLM-5.3 during the promotion, and one-ninth as much at list price. The comparison covers API billing only. It says nothing about how many tokens a model needs to complete the same job or how long the response takes.
Artificial Analysis exposes that missing part. Its independent model page gives GLM-5.3-Flash a score of 57.46 on Intelligence Index v4.1.1, close to the larger GLM-5.3 at 59.51. It calculates a cost of about $0.09 per Intelligence Index task using its own workload and price blend. Z.ai's launch article cites $0.045 per task at the discounted rate, so the two published figures should not be treated as interchangeable. They come from different calculations and snapshots.
The same independent run produced about 150 million output tokens across the index, compared with a 110-million median reported on the page. Artificial Analysis calls the model very verbose and its 48.7-token-per-second output notably slow. For an interactive coding assistant, that can mean more waiting even when each token is inexpensive. For an unattended batch job, verbosity can raise the bill and occupy a worker for longer. Teams evaluating the model need task-level cost and completion time, not just the rate card.
Why only 18 billion parameters are active
GLM-5.3-Flash's main efficiency claim rests on how little of the full model runs for each token. It has 320 billion parameters in total but activates 18 billion. Z.ai says the network has 45 layers, down from 92 in GLM-4.5, whose 355 billion total parameters include 32 billion active parameters. These figures come from the company's launch post, not an outside hardware audit.
The published architecture alternates three linear-attention layers with one sparse-attention layer. Linear attention carries local state without revisiting every prior token in the conventional way, while sparse attention retrieves selected global context. Z.ai's IndexPool component compresses four indexer key vectors into one before retrieval. The model configuration backs up that pattern across 45 layers and sets the maximum sequence length to 1,048,576 tokens.
According to Z.ai's measurements, the design cuts attention computation per head and layer by 3.0 times and average BF16 KV-cache size per layer by 4.4 times versus GLM-5.3. The company also concedes that its KV cache remains slightly larger than those of Kimi-K3 and DeepSeek-V4-Flash in the same comparison. That qualification matters: a one-million-token context limit is a capacity ceiling, while usable long-context performance depends on memory, serving software and retrieval accuracy.
The model is natively multimodal, with vision included in the base training rather than attached as a separate adapter. Z.ai says its pretraining corpus contains 30 trillion multimodal tokens. Its developer documentation accepts images as URL or Base64 content blocks and requires thinking mode to remain enabled. It recommends temperature: 1, top_p: 0.95 and maximum reasoning effort. Those defaults help explain why a quick swap into an existing client deserves a fresh latency and token-usage test.
The benchmark lead is narrower outside Z.ai's charts
Z.ai reports sizable gains over GLM-5.2 on two agent tests in its release results. GLM-5.3-Flash scored 63.4 against 46.2 on DeepSWE v1.1 and 48.8 against 26.2 on AutomationBench v1.0.6. Its internal Z.ai Code Bench result at maximum effort was 29.0, compared with 29.5 for Claude Opus 4.8. The company ran that coding test inside Claude Code 2.1.207, while its DeepSWE setup allowed a six-hour timeout and a 400,000-token context. These conditions make the scores more informative than bare percentages, but they remain vendor-run comparisons.
Artificial Analysis supplies a useful check without settling every question. Its current table places GLM-5.3-Flash at 57.46, just below Qwen3.8 2.4T A95B at 57.70. The index combines nine evaluations, including Terminal-Bench 2.1, SciCode, GPQA Diamond and Humanity's Last Exam. That is evidence that the smaller active model is competitive across the evaluator's mix. It does not prove parity on a team's repository, tool permissions or acceptance tests.
Developers can inspect and run the weights under MIT, but local deployment is still a data-centre job at this scale. The Hugging Face repository splits the model across 62 safetensor shards and publishes an FP8 build. Z.ai names four supported serving stacks, each with a separate recipe. An API trial is the easier first measurement for most teams; self-hosting becomes relevant where volume, data handling or hardware access can justify the operational work.
The Chinese-chip deployment is part of the release
Z.ai says in its launch account that the anonymous preview called ox-alpha ran on OpenCode and OpenRouter before the name was disclosed. The company describes it as the most popular model on those services for the week, though it does not publish request totals. More unusually, Z.ai says all preview traffic was served on a large cluster of domestically developed Chinese AI accelerators.
The serving stack is built on SGLang and separates multimodal encoding, prompt prefill and token-by-token decoding into independently scheduled worker pools. Z.ai also lists W8A8 quantization, mixed INT8, FP8 and BF16 cache formats, tensor parallelism and layer splitting among its deployment techniques. It claims a threefold improvement in end-to-end serving performance over its initial baseline on the same hardware, reaching per-token cost comparable with mainstream Nvidia GPUs. No chip model, cluster size, wattage or absolute throughput accompanies that comparison, so it cannot support a direct hardware efficiency ranking.
That omission limits the hardware claim. The downloadable model card still gives developers a concrete set of supported runtimes, while Z.ai gets to demonstrate that its public API does not depend entirely on Nvidia hardware. The model's independent output speed suggests that the stack still has room to improve at the user-facing layer.
The next useful evidence will come from repeatable repository-level trials: elapsed time, accepted patches, output tokens and peak memory under the same agent harness. Watch the API rates again after September 9, and watch whether independent providers can lift throughput above 48.7 tokens per second without changing quality. Those measurements will show whether GLM-5.3-Flash is a cheap default for sustained agent work or a capable model whose low token price hides a slower billable path to completion.