Daily AI news, sourcedVerdict
Unconfirmed. Per an expert-call note circulating in investor communities, Zhipu's next-generation base model has 2.5 to 3.5 trillion parameters, well above the 1.6 trillion figure that circulated earlier. The note says pre-training of the more aggressive architecture stalled for a while with the loss no longer falling, that the problem has likely been fixed and training should finish soon, and that the model is expected to reach Claude Fable-level flagship performance. A second X account relaying the same details adds that Zhipu plans to keep using a single base for six months or more with continuous post-training, and that recursive self-improvement (RSI) has made some progress but is not yet the main driver of capability gains. Zhipu has not disclosed the parameter count, the training stall or a Fable comparison. Background: on Aug 31, chief scientist Tang Jie said the next base will be larger, without giving a number, and that it should be usable at full scale on day one; he also said the GLM-5 base was planned to be used for more than six months. The earlier 1.6 trillion figure was a media extrapolation at the time, not a leak.
- Unconfirmed
Repeats the 2.5T-3.5T parameter range, the loss plateau and Fable-level target without naming a source; adds that a GLM-5 base is used for 6+ months with continuous post-training, and that RSI has made progress but is not yet the main driver. Not an official source.
Source: X post by @MaxForAI - Unconfirmed
Cites an expert-call note circulating in investor communities: next base 2.5T-3.5T parameters rather than the rumored 1.6T; pre-training hit a loss plateau under a more aggressive architecture, likely now fixed, training expected to finish soon; expected Claude Fable-level performance; inference speed and cost treated as hard constraints so it can be served at scale on day one. Not an official source.
Source: X post by @0xLogicrw Background, not confirmation: at Zhipu's Aug 31 interim results, Tang Jie said the next base will be larger but did not disclose parameters, aims for lower inference cost and full-scale serving on day one, and said the 744B GLM-5 base was planned to be used for more than six months.
Source: Wallstreetcn
- Is GLM-5.5 base model out yet?
- No. Zhipu (Z.AI) has not announced it as of Oct 21.
- When is GLM-5.5 base model expected?
- No official release date as of Oct 21.
- Unconfirmed
- Circulating with at least one public source we can link to; no official confirmation.
- Reported
- Reported by a news outlet; not confirmed by the company.
- Confirmed
- Confirmed by the company or another official source.
- Debunked
- Denied by the company, retracted by the original source, or shown to be false by the evidence.
We record what sources show, with dates. We don't predict or score how likely a rumor is.