GLM-6.0 Leak: Z.ai's Next Flagship Is 'Full Self-Training' — What We Know So Far

On the evening of August 31, 2026, Z.ai (Zhipu) held its first interim earnings call since listing on the Hong Kong exchange in January — and founder Tang Jie (唐杰) used part of it to define the company's next base model, naming it on the record for the first time: GLM-6.0. He positioned the model around a technical route Z.ai calls "Full Self-Training" (完全自训练), which he said corresponds to what the West calls RSI — recursive self-improvement. GLM-6.0 has no release date, no model card, no API, and no independent benchmark. This is a what-we-know-so-far piece about the first official confirmation of the name and the direction, not a launch story.

What the earnings call actually said about GLM-6.0

The account of the call comes from 观点网 (Guandian), a Hong Kong financial outlet that covered it the same evening; the signal was surfaced on X by @dotey. Every GLM-6.0 detail below is per Guandian's report of Tang Jie's remarks, and none of it has been confirmed by a Z.ai announcement or a model card.

Per the report, Tang Jie defined GLM-6.0 as a "Full Self-Training" model. The core capability he described is self-purification: the model covering the entire training pipeline — pre-training, mid-training, and post-training — with the aim of fully autonomous training and self-evolution. The headline is not a new parameter count; it is a claim that the model will take over parts of its own training.

He was explicit about the hard part. The biggest challenge, he said, is not scaling the model up; it is whether the model can achieve self-judgment — deciding for itself when to stop training, and correcting its own errors. He called that the core difficulty of the route and a key direction for future research. He also said future model work would fold in ethics and social governance as dimensions of the design, not afterthoughts.

The same call carried Z.ai's broader framing. Tang Jie laid out a formula — AGI business value equals intelligence ceiling × token consumption scale — and argued Z.ai should "sell tasks, not models." On the data side he noted that domestic large-model training corpora generally sit in the 30–50 trillion token range, and that the returns from simply stacking parameters are diminishing. That data-wall argument is the economic motivation for self-training: if human-quality text stops scaling, the model has to generate and judge its own training signal.

What "Full Self-Training" (RSI) actually means

RSI — recursive self-improvement — is the family of ideas where a model improves the processes that make it smarter, closing the loop between training and trained. In its strongest form, the model writes or tunes parts of its own training pipeline, so capability compounds without a human engineer in every loop. Frontier labs have written essays about it; almost none have publicly attached a shipping flagship to it. Z.ai's earnings-call positioning does exactly that, and that is why the phrase matters even before any GLM-6.0 code exists.

The positioning did not come from nowhere. It was prefigured in the internal letter Tang Jie sent to staff on July 11, 2026 — 《巨浪已来》 ("The Giant Wave Has Come") — announcing the "Touch High" (摸高) plan: two years of strategic investment, forgoing short-term monetization, to go after the next AGI high ground. The letter named four core engines; one of them was fully self-training. Coverage of the letter described that engine in concrete terms: build a high-quality synthetic data factory, use AI-versus-AI self-play to generate knowledge where human data is running out, and — inside a safety sandbox — give the system the ability to rewrite its own code, so that evolution speed is no longer bounded by human engineers.

That is a strong statement, and it should be read with the appropriate caution: all of it is the company describing an aspiration. No lab has demonstrated full self-training at frontier scale, and the letter is a strategy document, not a benchmark.

The challenge Tang Jie named is the honest part

The most valuable thing in the call, for a reader trying to predict whether GLM-6.0 will matter, is what Tang Jie said is hard. Self-judgment — knowing when to stop training and correcting errors autonomously — is the load-bearing assumption of the whole bet. If the model cannot reliably judge its own outputs, then synthetic data and self-play just amplify noise, and "self-training" collapses into a very expensive data pipeline with a human backstop.

The framing is consistent across the two statements: the letter's engine treats compute as evolution fuel; the call's challenge is about the judgment that makes that fuel useful. He also stressed a sizing principle — avoiding a model too big to be usable — which matters because the self-training bet is not a pure scaling bet. That is a meaningful signal about what GLM-6.0 is not: not just "the biggest GLM-family model yet."

Where GLM-6.0 sits in the line

Z.ai reported that the GLM-family has run through six generations in eleven months, with the company's own measure of intelligence rising from 32 to 60. The current flagship is GLM-5.3, announced mid-August with the API opened August 19 and open weights confirmed August 28. It is a text-in/text-out MoE with 40B active parameters, a 1M-token context window, and an Artificial Analysis Intelligence Index of 60 — the current open-weights high — priced at $1.40 per million input tokens and $4.40 per million output. The smaller GLM-5.3-Flash shipped MIT-licensed, and Z.ai says it drew 62 trillion tokens in its first six days (a vendor figure, not an independent measurement).

The interesting question is what happened to GLM-5.5. For most of August, analyst reports carried by Reuters and CGTN had a GLM-5.5 flagship arriving "within August" — and it did not appear. The earnings call named the next generation GLM-6.0, skipping past 5.5 entirely. Whether 5.5 was renamed, folded into 6.0, or never existed as an official product is unconfirmed. What the call does establish is the official reading: the next-generation base model is GLM-6.0.

Also unconfirmed: any parameter count, architecture, pricing, or availability for GLM-6.0. Z.ai's stated principle of "not too big to be usable," combined with the self-training framing, suggests the interesting change may be training paradigm rather than scale — but that is inference, not information.

What's confirmed, what's not

• Confirmed, per Guandian's report of the August 31 earnings call: Tang Jie named GLM-6.0 as the next-generation base model; he positioned it as Full Self-Training; he described self-purification across the training pipeline and self-judgment as the central challenge; and Z.ai reported H1 2026 revenue of ¥954M, up 399.7% year on year, with the API/MaaS business at 86.5% of revenue and a net loss of ¥2.07B, narrowed 12.1%.

• Confirmed independently of the leak: GLM-5.3's status — AA Intelligence Index 60, $1.40/$4.40, 1M context, text-only, open weights — measured by Artificial Analysis, not by Z.ai.

• Not confirmed: that GLM-6.0 exists as a shipping product; any release date; parameters, architecture, or pricing; whether GLM-5.5 is skipped or renamed; any benchmark; any availability through any provider.

• A single-source caveat: the GLM-6.0 quotes rest on one outlet's report of the call. Z.ai has issued no announcement about GLM-6.0 of its own as of this writing.

What a developer should do now

Nothing in this signal should move a production decision today, because there is no GLM-6.0 to call. The right response to a roadmap confirmation is to keep the option open, not to re-platform.

The useful thing to note is what a self-training flagship would change if it ships. GLM-family pricing has held a steady $1.40/$4.40 per million tokens across its recent flagship generations, undercutting the frontier closed labs on price while trailing them on independent index scores. A GLM-6.0 that genuinely trains itself would test the open-weights ceiling in a way no single-model launch has — and the first place that question gets answered is whether it appears in a provider catalog at a sane price with real failover around it.

When and if GLM-6.0 appears in a catalog, the pricing rule on OrcaRouter is simple: provider list price passed through with zero markup, live the same day — no renegotiation and no second contract. That is the same rule GLM-5.3 already runs under on the platform, where the current flagship is served at $1.26 / $3.96 per million tokens (a 10% launch discount on the $1.40 / $4.40 list), all through one API for 200+ models. And an unproven model with a roadmap as bold as RSI is exactly the case for routing by outcomes rather than names: point a slice of traffic at the new model with automatic failover to a proven one — GLM-5.3, DeepSeek V4 Pro, or GPT-5.6 Sol — so a model that turns out to be over-hyped degrades to the fallback instead of shipping broken. When there is no GLM-6.0 to route yet, the equivalent preparation is having the routing rule written before the model exists.

What to watch next

• A release date. Z.ai's flagship cadence has been roughly two months between generations; the call gives no date, and a "full self-training" model may not follow a normal cadence at all.

• The first concrete object: a model card, a weights drop, or an API ID. The moment GLM-6.0 gets an API identifier, the rumor becomes testable.

• GLM-5.5's fate. If a GLM-5.5 appears first, the roadmap reading changes; if it stays silent, the "6.0 skipped the number" reading is the more likely story.

• Whether other labs answer. RSI is usually discussed in essays, not earnings calls; if a major lab publicly attaches self-training to a flagship, the framing stops being a Z.ai quirk and becomes a trend.

• The first independent score. The current open-weights high on the Artificial Analysis Intelligence Index is 60 (GLM-5.3, tied with Kimi K3). A GLM-6.0 that clears that on an independent board is the event that matters — everything before it is roadmap.

None of this is a launch. GLM-6.0 is a name and a direction, confirmed on the record at an earnings call, and nothing more concrete exists yet. But the direction is the story: a major open-weights lab has publicly bet its next flagship on the model learning to train itself, and named the hard part in advance — judgment, not scale. That is the kind of signal worth writing down before it has a release date, because the moment GLM-6.0 gets a model card, the question stops being whether Z.ai means it and starts being whether it works.