The flagship went open — and its license did the talking
GLM-5.3’s 753B weights were already public; the bespoke license and a post-training-only upgrade claim are the real story.
The flagship went open — and its license did the talking: GLM-5.3’s 753B weights were already public; the bespoke license and a post-training-only upgrade claim are the real story. Issue 3, week of 9 September 2026.
Week of September 7–9, 2026 · Issue 3
A quiet window after the heaviest fortnight of the year — no new model launches from the major Chinese labs in three days. The movement was all in the layer underneath: open-weights provenance, licenses, runtimes, and a pricing cliff tonight. (Also filed: one correction to Issue 2, found by this cycle’s continuity pass.)
The Week in One Sentence
In a three-day window with zero launches, the biggest story was what had already shipped unnoticed: GLM-5.3’s 753B open weights — public since August 28 under a bespoke license, claiming flagship gains from post-training alone — while the STQ1_0 runtime stayed stubbornly un-merged upstream and GLM-5.3-Flash’s 50% promo pricing dies tonight at midnight Beijing time.
The Big Story: GLM-5.3’s open weights were already here — with a different license
(An explicitly labelled update on carried coverage — this resolves Issue 2’s “GLM-5.3 open weights, rumored” watch item, and corrects it; see Issue 2’s correction.)
What happened. No announcement, no blog post — but zai-org/GLM-5.3 is public on Hugging Face, and has been for longer than we reported: the repo’s initial commit is dated August 28, with community evaluation pull requests arriving August 28–29 (repo itself created August 25). Issue 2 said “rumored, no Z.ai primary yet” — that was a discovery miss on our side, now corrected. The artifact: 753,329,940,480 parameters across 141 shards (~755 GB), FP8-native (the bulk stored as F8_E4M3), 474K downloads to date.
Under the hood — and what’s not there. The card is strikingly thin for a flagship: model-type tag glm_moe_dsa (MoE with DSA-family sparse attention — the same architectural convergence this publication has been tracking since Issue 1), but no architecture prose, no active-parameter count, no stated context length (footnotes reference evaluation runs up to 1M tokens). What the card does say is the interesting part: GLM-5.3 “shares its base model with GLM-5.2, with gains attributed entirely to post-training.” Same base, +23.1 points on SWE-Marathon (42.5 vs 19.4), +20.7 on DeepSWE v1.1 (66.9 vs 46.2), +7.2 on Terminal-Bench 2.1 (88.2 vs 81.0) — all vendor-run, in a comparator table1 that leans hard into security and long-horizon agentic suites (CyberGym 84.5, ExploitGym 105/130 at 2h/6h, ALE-CLI 28.5).
The license is the load-bearing fact. The license field reads glm-5.3 — bespoke, tag license:other — not MIT, which is what GLM-5.3-Flash ships under. We read the field (verbatim, from the artifact); we have not yet summarized the terms — that’s an open ledger item, and until someone reads the LICENSE file, treat commercial use as unassessed. One vendor, two tiers, two licensing regimes, three days apart: Flash as the MIT-licensed on-ramp (weights August 25), the flagship as bespoke terms (initial commit August 28).
Why this matters. Three weeks ago the open-weights cascade was three models in five days; this week it’s quiet enough to see the texture of how it’s happening: no announcement for a 753B flagship drop (weights-first, docs-later — the Qwen Flash-Next pattern again), post-training claimed as the entire source of frontier movement (which matches what Goldman’s analysts and Tencent’s supervised-loop disclosure were saying from the outside), and per-artifact licensing now confirmed inside a single vendor at two model tiers.
What I would test. The post-training-only claim is testable at the margins: GLM-5.2 and GLM-5.3 should be near-identical on knowledge/recall tasks (same base) and diverge on long-horizon agentic and security suites (post-training). If 5.3 materially beats 5.2 on a knowledge benchmark, the claim softens. Weights for both are public — the experiment costs only compute.
Elsewhere in AI Engineering
GLM-5.3-Flash’s intro pricing dies tonight — Worth Watching
What happened. Re-read from the pricing page today: promo pricing still live — $0.075/M input, $0.25/M output, $0.015/M cache-hit — with the stated end: “24:00 on September 9, 2026 (UTC+8, Singapore time).” After that, list prices ($0.15/$0.50/$0.03) apply. GLM-5.3/5.2 API pricing is unchanged at $1.4/$4.4.
Why it matters. The Ox Alpha usage record (~23T tokens in six days, per Issue 1) was set while the model was free-to-cheap. Tonight is the cleanest price-elasticity experiment yet on a record-breaking launch: watch OpenRouter usage share over the next 48 hours. Even at list, Flash output remains ~17.6× cheaper than GLM-5.3 — the question is the shape of the curve, not the cliff.
The STQ1_0 afterlife: still patch-only — and one number confirmed — Worth Reading
What happened. Two carried items resolved. First, the llama.cpp PR queue (checked today) shows STQ1_0 support not merged upstream: the original NEON implementation (#22836) has been open since May 8 — the ternary format is a community format that predates Tencent’s Hy4 build — with an AVX2 kernel (Aug 19) and a Metal exclusion (Aug 31, draft) still open, and two CUDA PRs closed unmerged. No “hyv4” architecture support in the queue at all. Second: Terminal-Bench 2.1 for Hy4 preview is confirmed from the primary card at 85.4 — the number we flagged as single-secondary in Issue 2 is accurate; the card also adds rows we hadn’t captured (Apex Agents 37.1, Toolathlon Verified 74.1, HLE 43.4/55.4, Skillsbench 62.9).
Why it matters. Issue 2’s “the moment the patches merge, the 214-GiB Hy4 becomes a standard artifact” remains pending — and the queue suggests the merge is a format-war negotiation, not paperwork: the community’s STQ1_0 predates and differs from Tencent’s variant, so upstreaming means reconciling two implementations of a ternary format while Metal support is still being excluded. The 214-GiB artifact stays a build project.
GLM-5.3-Flash’s model card got an in-window edit — hybrid attention, now on the record — Worth Reading
What happened. The card was modified September 7, 12:13 UTC — the only in-window edit on any major card this cycle. The architecture section now states, verbatim: “For the first time in the GLM series, we introduce a hybrid architecture combining sparse and linear attention,” alongside Manifold-Constrained Hyper-Connections, a “newly trained base model,” and the 30T-token multimodal corpus. Downloads stand at 826,875 (vs ~346K in the first days). MIT license unchanged.
Why it matters. The hybrid sparse+linear claim was launch-materials-only when we covered it in Issue 1 — it is now model-card primary (P-V in our ledger). The mechanism we cannot yet verify from any Z.ai primary is unchanged too: IndexPool/KPool appears nowhere on the card, and NVIDIA’s docs name it differently (“KPool-indexed sparse attention”). The carried item survives a fourth consecutive check.
Research Radar
No paper from this three-day window met the read-beyond-abstract bar — the arXiv sweep could not be completed this cycle (our search backend is rate-limited until September 13), and nothing surfaced via the registry endpoints. Rather than force a weak entry, the Radar returns Sunday with a double sweep of the backlog.
Repo Radar
zai-org/GLM-5.3 (Hugging Face)
- What’s actually in it: 141 FP8-native shards (~755 GB), transformers config for the
glm_moe_dsamodel type, and an actively maintained evaluation section — community PRs keep adding results (HLE, DeepSWE, Toolathlon-Verified). Inspection depth: card + commit history; LICENSE file not opened. - What you can run: the weights via transformers; no serving recipes or quickstart visible at card depth.
- What’s missing: architecture prose, active-parameter count, context spec, hardware guidance — and a license summary. For a flagship artifact, the documentation gap vs. Tencent’s Hy4 release is itself a datapoint.
- Why open it: the raw material for the post-training-only experiment above — both 5.2 and 5.3 weights public, same claimed base.
My Read on the Week
(Synthesis and interpretation — not reported fact.)
1. The license map has fully bifurcated — inside vendors, not just between them. Apache-2.0 (Hy4) vs bespoke-permissive (Kimi K3) was an inter-vendor story in Issue 1; GLM makes it intra-vendor: MIT on Flash, glm-5.3 bespoke on the flagship, three days apart. Qwen did the same in the same fortnight (qwen-community-1.0 on Flash-Next, Apache-2.0 on Qwen-Drive). The pattern reads as deliberate tiering — permissive on-ramps, bespoke terms on the flagships — which makes “read the license field of the specific artifact” the only workable rule, now with four confirmations in three weeks.
2. Post-training is being positioned as the product. GLM’s own card (“same base as 5.2, gains entirely post-training”) is the most explicit vendor statement yet of what Goldman’s “ecosystem-driven training” and Tencent’s supervised loop implied from the outside: base models are freezing into platforms, and the frontier claim moves to the training loop. If the knowledge-benchmark experiment above holds the line (5.2 ≈ 5.3 on recall), the strategic read is that labs are competing on loops, not bases — and open weights are the distribution channel for the loop’s output, not the loop itself.
3. A quiet window is a portmanteau of the last one. Three days, zero launches — but the compounding assets all moved: a flagship’s weights slipped out unannounced, a runtime format fought its way through a PR queue, a pricing cliff got dated to the minute, and a card edit quietly upgraded a launch claim to documented. The burst cadence of launches makes easy news; the steady layer underneath is where an engineer should be reading.
What I’d Watch Next
- 48 hours after tonight’s promo flip. OpenRouter share for GLM-5.3-Flash before vs. after list pricing — the first elasticity datapoint for a record-setting launch (experiment protocol below).
- The GLM-5.3 LICENSE file. Someone will read it this week; the interesting question is whether it mirrors Kimi K3’s permissive-with-thresholds template or breaks new ground. We’ll summarize it next issue.
- Sunday’s double sweep. Our search backend returns September 13 — the deferred items (Zhipu’s chip claim, the Qwen4 rumor, GLM-5.5, the arXiv backlog, Tencent’s EVIE/AuK updates) all get their primary check then.
One Thing I’d Experiment With This Week
Run the price-elasticity measurement yourself, tonight. The promo ends at 24:00 UTC+8 — that’s a scheduled natural experiment on a model that processed a record ~23T tokens in six days.
- Setup: snapshot OpenRouter’s model rankings (rank, providers, per-token prices, and any usage-share figures visible) for GLM-5.3-Flash and its two nearest price neighbors tonight, before 24:00 Beijing time; repeat at +24h and +48h.
- Conditions: also snapshot one stable reference model (e.g., a DeepSeek V4 variant) to control for platform-wide drift.
- Measure: rank/share change through the flip; provider count change (providers re-price at different speeds — some may hold promo prices as loss leaders); cache-hit pricing visibility. Decision output: if share holds within ~20% at 2× price, agent workloads are pricing-inelastic below ~$0.50/M output and the next Flash-class launch has pricing headroom; if share halves, the free-ride era was the product, and the usage record tells us about price, not preference.
All vendor-reported figures are labeled as such; if independent evidence doesn’t exist yet, that’s stated rather than papered over. This issue’s claims ledger lives at research/2026-09-09/claim-ledger.md.
Footnotes
The GLM-5.3 card’s comparator columns include Opus 4.8, Claude Fable 5, and GPT-5.6 Sol — different comparator generations than Issue 1’s tables (Opus 4.6 Max / Opus 5). Cross-issue score comparisons across comparator generations are not apples-to-apples.↩︎
Citation
@online{dimsum_editorial2026,
author = {dimsum Editorial, the},
title = {The Flagship Went Open — and Its License Did the Talking},
date = {2026-09-09},
url = {https://news.llm101.in/issues/2026-09-09/},
langid = {en}
}