item  /  glm-5.3-flash  /  z.ai  /  mit weights

Z.ai Named It Flash. It Ranks 44th in Speed.

57 on the intelligence index, 50.2 tokens a second, and a rate that ends September 9.

Z.ai released GLM-5.3-Flash this morning: 320 billion total parameters, 18 billion active, a 1,048,576 token context window, image and video input, and open weights on Hugging Face under an MIT license. Artificial Analysis put it at 57 on the Intelligence Index, third out of 108 models in its class. That number is independent and it is real.

The number the launch post leads with is a different one. It plots the model at 0.045 per Intelligence Index task and marks it, in parentheses, as a reduced rate. Artificial Analysis's own page for the same model lists 0.09. The gap is not an error. Z.ai is running a 50 percent promotional rate that ends at 24:00 on September 9, 2026, Singapore time, and the frontier point in the chart is plotted at the promotional rate.

Which is fine, as long as you read the tag the way you read a shelf label. The big number is a rate. The number that hits your invoice is a rate multiplied by how many tokens the model emits, and this model emits a lot of them, slowly, with reasoning you are not permitted to switch off.

GLM-5.3-Flash is the first natively multimodal model in the GLM-5 series. It starts from a newly trained base model on a 30 trillion token multimodal corpus, and it is the first GLM to combine linear and sparse attention in one stack. Per the vLLM recipe, the 45 layer language model interleaves KDA linear-attention layers with NoPE sparse MLA layers and routes each token through 8 of 288 experts.

Linear attention carries local dependency through state modeling. Sparse attention pulls back the globally relevant context through a lightweight indexer. At a million tokens the indexer itself becomes the bottleneck, so Z.ai added IndexPool, which compresses four indexer key vectors into one by weighted pooling. Against GLM-5.3, the company reports 3.01 times less attention compute and a 4.44 times smaller KV cache, measured per head per layer and per layer in BF16 respectively.

Against the GLM-4.5 generation the shape is blunter. Similar total parameter count (320B against 355B), but roughly half the activated parameters (18B against 32B) and half the layers (45 against 92). The model also adopts Manifold-Constrained Hyper-Connections, which Z.ai describes as a scaling efficiency improvement. There is no technical report for any of this yet. The Hugging Face card cites the February GLM-5 paper, which predates every one of these changes.

It spent its first week on OpenCode and OpenRouter under the name ox-alpha, unlabelled, and Z.ai says it became the most used model of the week there. All of that traffic ran on Chinese AI chips.

Z.ai's pricing page prints the list rates struck through next to the current ones: 0.15 input, 0.03 cached input, 0.50 output, halved to 0.075, 0.015 and 0.25 for the promotional window. The claim that it beats GLM-5.2 at roughly one tenth the price checks out against list, not against the promotional column. GLM-5.2 and GLM-5.3 both sit at 1.4 input and 4.4 output, so 0.15 over 1.4 is 10.7 percent and 0.50 over 4.4 is 11.4 percent. Roughly one tenth is honest.

What is worth putting in your calendar is the second week of September. Anything you size, budget or benchmark this fortnight is priced at a rate that ends on the 9th. If you are building a cost model for an agent that runs into Q4, build it at 0.15 and 0.50.

The genuinely durable line on that table is cached input at one fifth of input. For a coding agent replaying a large stable prefix on every turn, that ratio moves more money than the headline rate does.

Artificial Analysis publishes a verbosity figure alongside its intelligence score: the total output tokens a model burns to complete the whole Intelligence Index. GLM-5.3-Flash used 150 million against a class median of 110 million, ranking 32nd of 108 on that measure. Running the index cost 138.02 dollars in total at list rates.

You cannot dial that down. Z.ai's own documentation states that thinking.type supports enabled and nothing else, so reasoning cannot be disabled, and the recommended settings are temperature 1, top_p 0.95 and reasoning_effort at max. Reasoning tokens bill at the output rate. A per-token price advantage that gets partially spent on tokens you did not ask for is still an advantage, but it is a smaller one than the tag suggests.

The speed number is the one that earns the title. Artificial Analysis measures 50.2 output tokens per second against a class median of 67.1, which places it 44th out of 108 comparable models. Time to first token is genuinely good at 1.47 seconds against a 2.16 median, so it feels responsive and then finishes slowly. For an interactive tool that combination reads well. For a long-horizon agent chewing through a million token context, wall clock is the thing you are actually paying for.

One caveat in Z.ai's favour: that 50.2 is measured against the Z.ai endpoint, not the model. The next figure is where that distinction gets expensive.

The launch post's benchmark table carries six columns. The prose names two of them: GLM-5.2, which it beats everywhere the older model has a number, and Claude Opus 4.8, which it says it approaches. Both readings hold. Counting the fourteen rows myself, it takes 9 and loses 5 against Opus 4.8, including a 13 point gap on NL2Repo and a 0.5 point gap on CharXiv Reasoning.

Two columns in that same table are never mentioned in the text. Gemini 3.7 Flash appears in nine rows and wins six of them, taking Terminal-Bench 2.1, DeepSWE, AutomationBench, BabyVision, MVbench and MMVU. GPT-5.6 Terra appears in ten and wins four, including both of the top-line coding rows. Neither name appears anywhere in the post outside the table.

That is not a scandal, it is a positioning choice, and it is a fair one given the price gap. But if you are picking a low-cost multimodal agent model this week, the model you are actually choosing between is the other Flash, and Z.ai's own table says that fight is not close on video and image understanding.

MIT weights are the part of this release that will still matter in October. The practical envelope, from the vLLM recipe: the default FP8 checkpoint is roughly 306 GiB of weights before runtime and KV cache, the BF16 variant is about twice that, and the current vLLM path supports NVIDIA Hopper and newer, with FlashInfer 0.6.17 or later required for NoPE sparse MLA. That puts a clean single-node deployment at an eight GPU box, or one GB200 tray at tensor parallel four.

One detail in that recipe interacts directly with the headline efficiency claim. Hopper does not support an FP8 KV cache for this model and has to run BF16 KV. The 4.44 times cache reduction Z.ai advertises is measured in BF16 per layer, so it still applies, but the additional halving you would get from an FP8 cache is available on Blackwell and not on H100 or H200. If you sized a Hopper cluster off the marketing number, resize it.

SGLang, TokenSpeed and KTransformers are also listed, and Z.ai has published an SGLang cookbook entry. Twenty one community quantizations were already on the Hugging Face model tree the day it landed.

Everything above is a caveat about the tag, so here is the other side of the ledger, because most of the model is very good.

The 57 on the Intelligence Index was measured by a third party across nine evaluations, and third of 108 in class is not a marketing artefact. Time to first token beats the class median by a wide margin. Cached input at a fifth of input is a durable structural rate, not a promotion. GDPval-AA v2 at 1773 is the highest number in Z.ai's entire table, ahead of Opus 4.8 at 1582 and Gemini 3.7 Flash at 1527, and that evaluation was run by Artificial Analysis rather than by Z.ai. The weights are MIT, which is the least restrictive license anyone has attached to a model this capable.

The base model comparison is more honest than most launch posts bother to be. Z.ai prints GLM-5.3-Flash-Base next to GLM-5-Base and loses four of the five listed benchmarks: MMLU 88.1 against 88.3, BBH 86.6 against 87.4, HellaSwag 87.1 against 88.1, SimpleQA 33.5 against 36.0. It wins LiveCodeBench-Base 37.6 against 34.4. Doing that with 18 billion active parameters against 40 billion is the actual efficiency result, and the post describes it accurately as remaining competitive rather than claiming a sweep.

One claim I could not check at all is the serving story. Z.ai reports a 3 times end-to-end improvement over its own baseline on Chinese AI chips, reaching hardware efficiency and per-token cost comparable to mainstream NVIDIA GPUs. The chip vendor is not named, the model of accelerator is not named, and no absolute figure is published on either side of that comparison. The 3 times is measured against Z.ai's own starting point on the same hardware, which is a real engineering result and not a comparison to anything from NVIDIA.

researchaudio.io

Sources: Z.ai launch post  ·  Z.ai model documentation  ·  Artificial Analysis  ·  OpenRouter provider board  ·  Hugging Face model card  ·  TestingCatalog