← Gautam Parab

Z.ai Named the Kernel Flag. It Did Not Name the Chip.

There is a way to read a vendor engineering post that has nothing to do with whether you trust the vendor. Sort its claims by how much work it would take someone outside the company to check them, then see where the headline sits on that list. Usually it sits at the bottom.

Z.ai published one today, β€œToward Recursive Self-Improvement: How GLM Built Its Own Inference Infrastructure”, and it is a better piece of writing than that framing suggests. It is specific about failure, names the library version behind a concurrency bug, links a merged upstream patch, and ends by disclaiming the thing in its own title. I went in expecting to take it apart and came out thinking the problem is structural rather than rhetorical. The parts of this post that can be audited are the small ones. The parts that carry the argument cannot be audited at all.

The headline claims are four. A production inference service built β€œfrom scratch on a cluster of more than 100,000 Chinese-made AI accelerators,” on which β€œall production inference for GLM-5.3-Flash runs.” Roughly 3Γ— end-to-end serving improvement. β€œBoth hardware utilization efficiency and per-token cost reached levels comparable to mainstream NVIDIA GPUs.” And a launch that went β€œfrom initial model adaptation to production readiness in less than two weeks,” much of the work β€œcarried out by an Infra Agent powered by GLM-5.3.”

The eight claims in Z.ai's inference-infrastructure post, positioned by how load-bearing each is and how independently checkable A two-axis chart. The horizontal axis runs from opaque on the left to checkable on the right; the vertical axis runs from incidental at the bottom to load-bearing at the top. Four claims sit in the upper-left region, load-bearing but not checkable: more than 100,000 accelerators, per-token cost parity with NVIDIA, a 3x end-to-end throughput gain, and the assertion that an Infra Agent did work that would have taken a team of engineers weeks. Two claims sit in the middle of the right-hand side, secondary but checkable against public dashboards: 62 trillion tokens in six days, and being the most-used model on both platforms. Two claims sit in the lower right, incidental but fully checkable in public code: a DeepEP version 1.2.1 GIL bug, and a tf32x3 precision fix merged as Flash Linear Attention pull request 1180. The upper-right region, where a claim would be both load-bearing and checkable, holds nothing. CLAIM CHECKABILITY Β· Z.AI INFERENCE POST Β· 17 SEP 2026 OPAQUE CHECKABLE LOAD-BEARING INCIDENTAL More than 100,000 Chinese-made AI accelerators β€” no vendor named β€” Z.ai, 17 Sep 2026 Per-token cost and utilization comparable to mainstream NVIDIA GPUs β€” no accuracy number given β€” Z.ai, 17 Sep 2026 Roughly 3x end-to-end serving improvement, measured against the project's own initial baseline β€” Z.ai, 17 Sep 2026 An Infra Agent did work that would have taken experienced engineers weeks β€” no control condition β€” Z.ai, 17 Sep 2026 More than 62 trillion tokens in six days β€” checkable against OpenRouter and OpenCode dashboards β€” Z.ai, 17 Sep 2026 Most-used model on both OpenCode and OpenRouter within a week of launch β€” Z.ai, 17 Sep 2026 DeepEP v1.2.1 intranode dispatch and combine did not release the Python GIL β€” readable in public source β€” Z.ai, 17 Sep 2026 tf32x3 precision fix merged as Flash Linear Attention PR #1180 on 27 Aug 2026 β€” verified merged 100,000 accelerators Per-token cost parity with NVIDIA 3x end-to-end throughput Infra Agent saved a team weeks 62T tokens in six days Most-used, both platforms DeepEP v1.2.1 GIL bug tf32x3 fix (PR #1180) Placement is my reading of the post, not a measurement; the two lower-right claims I confirmed directly in public code.
The upper-right quadrant is empty. Nothing this post rests its case on can be checked from outside the company, and everything that can be checked is a detail.

The chip is never named

I searched the full rendered text for every Chinese accelerator vendor I could think of: Huawei, Ascend, Cambricon, Biren, Moore Threads, Hygon, Kunlun, Enflame, MetaX, Iluvatar. Zero hits. The phrase β€œChinese-made AI accelerators” appears once, a close paraphrase without the β€œAI” appears once more, and after that it is just β€œaccelerators.” NVIDIA is named once, as the comparison. There is no chip model, no vendor, no process node, no interconnect.

That omission is not a detail, because the post itself makes memory the whole story: β€œWe faced relatively limited chip memory capacity and bandwidth, while also needing to support a new model architecture, a 1M-token context window, and multimodal requests.” Every optimization it then describes (trading compute for bandwidth, trading communication for device memory, W8A8, cache quantization, Layer Split) is a response to a specific memory-capacity and bandwidth budget. Without knowing the part, a reader cannot tell whether those were heroic workarounds or routine tuning, and cannot check the efficiency claim against anything.

Is a 100,000-accelerator domestic cluster plausible? At the level of national output, yes. Cambricon reported a 108% jump in first-half revenue to about 6 billion yuan in its half-year filing on 7 August, and Biren has said first-half revenue could rise by up to 2,107%, which is company guidance rather than audited results. One operator holding 100,000 parts is arithmetically unremarkable against that. But it is a statement about the country, not about this cluster: with no vendor named, the claim cannot be checked against anyone’s shipments. The number does not look wrong. It has been placed permanently out of reach.

A cost number with no quality number

Per-token cost is not a property of hardware. It absorbs electricity, capital cost, amortization schedule, and utilization. I spent an earlier piece on how much the depreciation assumption alone can move that figure. Two operators running identical silicon can report per-token costs that differ by more than the gap Z.ai is claiming to have closed.

What sits next to the cost claim is nothing. The paragraph reporting utilization and cost parity is the same one that lists the serving stack: W8A8 quantization, β€œmixed-precision cache quantization using INT8/FP8/BF16,” Layer Split, an Encode-Prefill-Decode disaggregated architecture. It carries no accuracy number, and neither does any other paragraph in the post. A cost per token is comparable to another only if both serve the same quality of output, and quantization is the knob that trades one for the other.

The 2026 literature on that trade is clear that the answer is configuration-specific and sharply non-linear. Lee and colleagues measured a β€œquality cliff”: Llama-3.1-8B-Instruct is statistically indistinguishable from full-precision KV down to 2.322 code bits per value and then collapses at 2.0 bits. Xu, Kumarappan and Zhou found something worse for anyone relying on aggregate metrics. Across eleven instruction-tuned models, Mistral-7B loses 15.2% of its refusals at 1.03Γ— perplexity, with β€œno universal safe bit-width” and model-specific transitions they describe as invisible to standard measurement. A fourth group auditing INT8 and INT4 caches in retrieval settings found INT8 near-lossless, but at INT4, among answers that stayed factually correct, over 90% of the faithfulness changes were negative. The accuracy metric does not see it.

Three 2026 measurements of what cache quantization costs, and why the usual metrics miss it Three panels. The first reports that Llama-3.1-8B-Instruct stays statistically indistinguishable from full-precision KV cache down to 2.322 bits per value and collapses at 2.0 bits, a narrow quality cliff. The second reports that Mistral-7B loses 15.2 percent of its refusals at only 1.03 times perplexity, so a three percent change in the aggregate metric conceals a fifteen percent change in behaviour. The third reports that INT8 caches are near-lossless, while at INT4 more than 90 percent of faithfulness changes are negative even among answers that remained factually correct. All three findings concern models other than GLM. WHAT CACHE QUANTIZATION COSTS Β· THREE 2026 RESULTS Β· OTHER MODELS KV BITS / VALUE PERPLEXITY VS REFUSALS INT8 VS INT4 2.0 15.2% >90% bits where Llama-3.1-8B collapses; 2.322 is still indistinguishable from FP16 of Mistral-7B's refusals lost at 1.03x perplexity - a 3% change hides a 15% one of faithfulness changes at INT4 are negative, among answers that stayed correct 2.0 2.322 perplexity refusals 1.03x perplexity, about a 3% change β€” Xu, Kumarappan and Zhou, 1 Jun 2026 15.2% of refusals lost β€” Mistral-7B β€” Xu, Kumarappan and Zhou, 1 Jun 2026 +3% -15.2% INT8 near-lossless INT4 accuracy and faithfulness fall Each panel measures a different model; none tested GLM. The shared finding is that aggregate metrics understate the loss.
The precisions Z.ai names sit on the safe side of every cliff these papers found. That is reassuring, and it is still not a measurement of Z.ai's stack.

Read fairly, that literature mostly exonerates Z.ai’s choices. The cliffs all sit at or below INT4, well under the INT8 floor of the cache configuration Z.ai names, and a January benchmarking study of microscaling formats found the same shape at the same boundary: MXFP8 β€œconsistently achieves near-lossless performance,” while MXFP4 β€œintroduces substantial accuracy degradation.” None of the four papers tests W8A8 directly, and none of the three cache studies tests a GLM. My guess is that this serving stack costs very little quality. But that is my inference from other people’s Llamas and Qwens, not a measurement of anyone’s GLM, and it is not what the post asked me to take on faith.

It would have been easy to close. Z.ai’s own model card for GLM-5.3-Flash is scrupulous about every other reproduction detail. It specifies the reasoning_effort level to use for leaderboard runs, gives DeepSWE a temperature of 0.95, a top_p of 1.0, a six-hour timeout and a 400K context, and documents sampling parameters to two decimal places. About numerical precision it says nothing whatsoever: the words quantization, W8A8, INT8 and FP8 do not appear on it. So the published benchmark scores and the published cost claim cannot be confirmed to describe the same machine, and the gap is in the one field the infrastructure post says production changes.

The 3Γ— is against itself

The post is honest about this if you read carefully: throughput tripled β€œrelative to the initial baseline,” which is the system’s own first working run on new hardware. That is a respectable engineering result, and not a comparison to anything outside the project. A 3Γ— gain over your own first attempt tells a reader how rough the first attempt was, which the post says plainly: β€œthe ecosystem was immature, kernel support was incomplete, and much of what should have been documented had to be guessed.”

The one claim I could test

Which brings me to the only headline number with a public scoreboard behind it. GLM-5.3-Flash was trialled anonymously as Ox-Alpha, and the post says that β€œwithin a week of launch, it became the most-used model on both platforms, processing more than 62 trillion tokens in six days.”

OpenRouter and OpenCode both publish usage. I pulled both on 17 September. OpenRouter’s rankings, computed from usage through 16 September, put GLM 5.3 Flash third over the trailing seven days at 11.3 trillion tokens, behind GPT-5.6 Luna at 17.3 trillion and Tencent’s Hy4 preview at 12 trillion. OpenCode’s page for the model ranks it fifth in the last week with an 8.6% share, and gives a cumulative 56 trillion tokens over a window it labels 24 July to 17 September.

The windows differ, so the only fair comparison is a rate, and the arithmetic is mine. 62 trillion over six days is about 10.3 trillion tokens a day. OpenCode’s 56 trillion across those 56 days averages 1.0 trillion a day. OpenRouter’s current seven days run about 1.6 trillion a day. The claimed launch-week rate is roughly ten times the longer-run average on one platform and about six times the model’s present rate on the other.

Implied daily token rate: Z.ai's launch-week claim against the two public dashboards Three rows, each a horizontal stem ending in a dot, showing trillions of tokens per day. Z.ai's launch-week claim of more than 62 trillion tokens over six days implies about 10.3 trillion per day, by far the longest stem. OpenRouter's trailing seven days, from usage data through 16 September 2026, shows 11.3 trillion tokens, about 1.6 trillion per day. OpenCode's cumulative 56 trillion tokens over the 56 days from 24 July to 17 September 2026 averages 1.0 trillion per day. The claimed rate is roughly six times the current OpenRouter rate and about ten times the longer-run OpenCode average. IMPLIED TOKENS PER DAY Β· GLM-5.3-FLASH / OX-ALPHA Β· SNAPSHOT 17 SEP 2026 About 10.3T tokens/day β€” more than 62T over six days β€” Z.ai, 17 Sep 2026 About 1.6T tokens/day β€” 11.3T over the trailing seven days β€” OpenRouter rankings, usage data through 16 Sep 2026 About 1.0T tokens/day β€” 56T over 56 days β€” OpenCode model data page, snapshot 17 Sep 2026 Z.ai's launch-week claim OpenRouter, trailing 7 days OpenCode, 24 Jul - 17 Sep 62T tokens over six days 11.3T tokens, data to 16 Sep 56T tokens over 56 days 10.3T / day 1.6T / day 1.0T / day Daily rates are my own arithmetic from each source's stated total and its stated window. OpenRouter excludes traffic kept private.
Different windows, so the only fair comparison is a rate. The claimed launch-week rate is not a thing either dashboard has ever shown.

That is not a refutation. A free stealth preview can draw traffic that a paid endpoint never sees, the 62 trillion may be a sum across both platforms, and OpenRouter states that it excludes requests which users or apps keep private. But neither public dashboard shows anything resembling that rate, and the β€œmost-used on both platforms” standing did not persist for even a month: today the model is third on one and fifth on the other.

A unit problem sits underneath all of this, and it cuts against the cost claim too. OpenCode reports that 93% of this model’s input tokens are served from cache, on sessions averaging 3.6 million tokens. A β€œtoken processed” at a 93% cache-hit rate is mostly a cache read, not a forward pass. Token volume is a soft unit for traffic, and a softer one for the serving work that a per-token cost is supposed to price. OpenRouter says as much in its own methodology note: the rankings β€œmeasure adoption, not quality,” and token volume β€œis also not a count of requests, users, or spend.”

The agent, and what would have made it a measurement

The recursive-self-improvement framing is the part I expected to object to, and the post mostly beats me to it: β€œChoosing objectives, setting boundaries, and assessing risk remain human responsibilities. We believe humans should continue to hold that line for a long time to come.” That is a more careful sentence than the title.

What is missing is a control. The only comparative claim about the Infra Agent is a counterfactual: the model completed a task β€œthat would previously have taken a team of experienced infrastructure engineers weeks,” with no task scope, no baseline team, and no holdout. The agent’s quantified results are the strongest material in the piece, because they are narrow: a division optimization that cut one kernel’s execution time by 9.6%, a tiling change worth a 1.71Γ— speedup over the previous version, and a KV-transfer overhead brought from above 20% down to below 1% against a stated 5% acceptance criterion.

And the most checkable claims in the whole post are also its most obscure. Z.ai says the agent found a numerical-accuracy bug in the KDA kernel’s context-parallelism path, where Triton’s tl.dot defaulted to TF32 even on FP32 inputs, and that the fix, input_precision="tf32x3", went upstream as pull request #1180 in Flash Linear Attention. That PR exists, is titled β€œ[CP] use tf32x3 affine chain in kcp,” was merged on 27 August, and its description matches the blog’s account: tf32x3 in the state update and affine-transform merging, β€œto avoid precision loss in long context.” It adds 140 lines and removes 13 across six files, 59 of them a context-parallel KDA test. The claim about DeepEP v1.2.1 is similarly precise: intranode_dispatch and intranode_combine failed to release the Python GIL while internode_dispatch already did, starving the transfer thread. Anyone can read that code.

So the post’s credibility gradient runs exactly backwards from its emphasis. A precision flag in a Triton kernel is fully auditable, down to the test that covers it. One hundred thousand accelerators are not.

The industry worked this out before, for databases. From 1985 through 1988, vendors used the TP1 and DebitCredit benchmarks, or their own readings of them, to, in the Transaction Processing Performance Council’s own telling, β€œmuddy the already murky performance waters.” Omri Serlin got eight companies to form the TPC on 10 August 1988, and the council’s answer was procedural: its first benchmark β€œspecified that all benchmark testing data should be publicly disclosed in a Full Disclosure Report.” The TPC then hit the harder half of the problem. Having built a good benchmark and a good review process, it β€œhad no means to control how those results were used once they were approved.” The gap between measured results and marketing claims got its own name: β€œbenchmarketing.” The Fair Use policies adopted in June 1991 asked for β€œfidelity, candor, and due diligence.”

Machine learning has a version of this now, and it reads strangely against today’s post. MLCommons’ MLPerf inference rules answer the question β€œwhat will make my submission more likely to be audited” with three conditions: performance β€œnot consistent with the known or expected characteristics of the hardware”; reviewers who β€œlack insight into how the measured performance was achieved”; and β€œhardware and software [that] is not reasonably available to the general public.” The Z.ai post satisfies all three at once. The same rules permit aggressive quantization, to β€œany numerical format that achieves the desired quality,” on two conditions: that the method β€œmust be publicly described at a level where it could be reproduced,” and that accuracy bounds are met. Z.ai published the method names and no accuracy bound.

None of which is an accusation. A company blog is not a benchmark submission and was never under those rules. That is the point: the only regimes that ever fixed this problem were procedural, and a launch post sits outside all of them by construction.

Three things would have made this a shorter essay. Name the accelerator. Put one accuracy number next to the cost-parity claim, measured on the stack that actually serves traffic. And state on the model card whether the published evaluations were run under the production serving configuration or under BF16. The engineering here looks good. Good enough that it does not need the parts nobody can check.

References

  1. Z.ai (2026). Toward Recursive Self-Improvement: How GLM Built Its Own Inference Infrastructure. 17 September 2026.
  2. Z.ai / zai-org (2026). GLM-5.3-Flash model card. Hugging Face. Accessed 17 September 2026.
  3. OpenRouter (2026). LLM Rankings. Usage data through 16 September 2026; snapshot taken 17 September 2026. Rankings data licensed CC BY 4.0.
  4. OpenRouter (2026). Ox Alpha β€” model page. Released 20 August 2026; snapshot taken 17 September 2026.
  5. OpenCode (2026). GLM-5.3-Flash (formerly ox-alpha) β€” usage, cost and rank. Window labelled 24 July–17 September 2026; snapshot taken 17 September 2026.
  6. fla-org (2026). [CP] use tf32x3 affine chain in kcp. Flash Linear Attention pull request #1180, merged 27 August 2026.
  7. Lee, Kim and Kim (2026). SemKV: Semantic Mixed-Precision KV Cache Quantization Guided by the Quality Cliff for Long-Context LLM Inference. arXiv:2608.28911v1, 28 August 2026.
  8. Xu, Kumarappan and Zhou (2026). Alignment Collapse Under KV Cache Quantization: Diagnosis and Mitigation. arXiv:2606.09864v2, 1 June 2026.
  9. Asad, Bilal, Ali, Haseeb and Hougen (2026). Faithfulness Is Not Free: Auditing Offline KV-Cache Quantization in Retrieval-Augmented Generation. arXiv:2608.30996v1, 31 August 2026.
  10. Benchmarking Post-Training Quantization of Large Language Models under Microscaling Floating Point Formats (2026). arXiv:2601.09555, 14 January 2026.
  11. MLCommons (2026). MLPerf Inference Rules. inference_policies repository; file last revised 20 August 2026.
  12. Transaction Processing Performance Council. A Brief History of the TPC. Accessed 17 September 2026.
  13. Cao, Ann (2026). Cambricon posts 108% surge in first-half revenue amid China’s massive AI chip drive. South China Morning Post, 7 August 2026.
  14. Cao, Ann (2026). AI chipmaker Biren projects up to 22-fold revenue surge amid China’s hi-tech boom. South China Morning Post, 17 August 2026. Company guidance, not audited results.