DeepSeek’s rumored 10T-parameter model could begin training in late 2026 to April 2027
zephyr_z9 · x · 2026-07-23
A speculative timeline post says DeepSeek’s rumored 10T-parameter model could start training between late 2026 and April 2027.
The reasoning assumes:
- 1/64 sparsity
- 150B active parameters
- a start date that would put it 12–15 months behind Anthropic
It is an estimate, not an announcement, but it signals how aggressively people are modeling the next frontier of large-model scaling.
More from Models
- Claude Opus 5 appears to be rolling out to some accounts — gaganghotra_ · 2026-07-23
- Grok 4.5 tops VulcanBench v3 with 91.3% on real repo engineering tasks — XFreeze · 2026-07-23
- Google AI Search reportedly forgets context within the same chat — Typical-Function8542 · 2026-07-23
- Kimi K3 tops Code Arena and reignites the debate over AI economics — Exponential View (Azeem Azhar) · 2026-07-23
- ChatGPT answers a question nobody asked with a full Katseye fan pitch — RANDOM_OVERTHINKER_ · 2026-07-23
- GLM-5.2 adds vision support and is now open source, with SGLang run instructions — baseten · 2026-07-23