Rumor: DeepSeek Training 2T-Parameter Model with 8T Planned, Entirely on Huawei Ascend Chips

From Sept 21–22, posts on X relayed claims that DeepSeek is training a 2T-parameter model with an 8T model planned, reportedly trained entirely on Huawei Ascend chips. None of this is officially confirmed, but if true it would mark a milestone for DeepSeek in both model scale and training hardware.

Confirmed

The rumor was relayed by multiple X users: Terminator857 and Hesamation both reported DeepSeek training a 2T-parameter model with an 8T model planned; for reference, DeepSeek's current Flash is 552B parameters, Pro (V4 Pro) is 1.6T total with 49B active weights per token, and Kimi K3 is 2.8T
Joseph Jacks, relaying @jenzhuscott, added that the 2T model is trained entirely on Huawei Ascend chips, calling it a turning point showing a top Chinese model lab can train frontier models outside the Nvidia ecosystem (no evidence attached)
Hesamation added that DeepSeek expects new Huawei training chips between Q4 2026 and Q1 2027, and its CEO told investors the shift to domestic chips "must succeed"

Unconfirmed

The specifics and progress of the 2T/8T models come from unverified rumors; DeepSeek has not responded
X user EvilDragon233 speculated DeepSeek V5 will be an 8T-total/250B-active MoE possibly not released until late 2027; teortaxesTex joked about DeepSeek emphasizing defense while scaling up

Why it matters

If the 2T model is real, it would be DeepSeek's largest model yet, well beyond the current 1.6T V4 Pro
A full shift to Huawei domestic chips for large-scale training would mean breaking Nvidia dependence, an industry-signaling move
teortaxesTex argued current LLM scaling is still alchemy: in theory equal active parameters should yield equal performance, Meta ran a dense 405B training on 16,000 H100s two years ago, and sparsity's MFU loss is manageable—implying the payoff of pure parameter scaling is debatable

2026-09-21 ~ 2026-09-22 · 6 related posts

Primary sources

DeepSeek reportedly training a 2T-parameter model, with an 8T model on the roadmap — Terminator857 · 2026-09-21 20:38 (UTC+08:00)
Report: DeepSeek training a 2T-param model with 8T planned, Huawei chips due late 2026 — Hesamation · 2026-09-21 19:16 (UTC+08:00)
Hesamation follows up on DeepSeek's 2T/8T training and domestic chip plans — Hesamation · 2026-09-21 19:16 (UTC+08:00)

[source] Report: DeepSeek training a 2T-param model with 8T planned, Huawei chips due late 2026 — Hesamation · 2026-09-21
[source] Hesamation follows up on DeepSeek's 2T/8T training and domestic chip plans — Hesamation · 2026-09-21
[source] DeepSeek reportedly training a 2T-parameter model, with an 8T model on the roadmap — Terminator857 · 2026-09-21
Unverified leak claims DeepSeek V5 is an 8T-param MoE with 250B active, due late 2027 — teortaxesTex · 2026-09-21
DeepSeek rumored to train at 8T tokens as X users slam LLM scaling as alchemy — teortaxesTex · 2026-09-21

1 near-duplicate retellings: JosephJacks_