Rumor: GLM-5.5 base model at 2.5-3.5T params, delayed by loss plateau
khademinori · x · 2026-10-21
According to a leak from @MaxForAI (unverified), key details on Z.ai's next-gen GLM-5.5:
- Scale: base model at 2.5T-3.5T parameters, notably larger than the previously rumored 1.6T; aggressive architecture changes aim to boost both training and inference speed and cut inference costs, with large-scale serving from day one
- Setback: base training previously hit a loss plateau, slowing progress below market expectations; the issue appears recently resolved
- Capability targets: expected to reach flagship level, continuing to push toward Claude-like usage feel, with heavy post-training investment in task environments
- Roadmap: one GLM-5 base serving 6+ months with continuous post-training; RSI has made progress but isn't yet a main capability driver
The quoting user's focus: whether GLM-5.5 can deliver 3T scale, flagship capability, inference speed and cost all at once — especially for long agent tasks requiring many consecutive calls.
More from Companies & People
- Google DeepMind London opens 2027 student researcher role in multi-agent coordination — edelwax · 2026-10-22
- OpenAI told Australian parliament its local subsidiary made just $3.87M — gaganghotra_ · 2026-10-22
- YC-backed Stepful trains 30,000 healthcare students a year, raises $55M Series C — ycombinator · 2026-10-22
- Microsoft launches Contact Center Intelligence to bring agentic AI to customer service — satyanadella · 2026-10-22
- WSJ profiles Demis Hassabis: solve intelligence first, then solve everything else — clmt · 2026-10-22
- Eval builder Ken joins Google DeepMind as forward deployed engineer to co-design benchmarks — osanseviero · 2026-10-22