A post on X at 04:12 UTC on September 26, 2026 announced a DeepSeek V5 leak under the headline Beats Astra. It says “DeepSeek is reportedly preparing an imminent V5 launch”, that Liang Wenfeng calls it the company’s biggest bet, that it is “Rumored at 2 trillion parameters (not 3T)”, and that it is the “first DeepSeek model to train fully on Huawei Ascend chips instead of Nvidia”. The image attached to the post is DeepSeek’s whale logo with v5 typed beside it. There is no chart, no table and no score in it.
Some of the reporting underneath is real. The conclusions drawn from it are not. This page separates the two, and it stays open until DeepSeek publishes something.
On this page · 10 sectionsOpen
- DeepSeek has said nothing about V5. Its news page and API change log end at V4.1 Flash, released September 10, 2026, and DeepSeek has said a V4.1 Pro comes next.
- The only V5 timing in print is a Bank of America forecast, relayed by 36Kr on September 2: V5 is expected to debut in the second half of 2026. A forecast, not a launch notice.
- The Information reported on September 23 that DeepSeek is training a model of about 2 trillion parameters, with 8 trillion as the next goal. No report calls that model V5.
- The biggest bet quote is about chips. The Information reported on September 21 that Liang Wenfeng told investors training on Huawei chips is one of the company’s biggest bets.
- Training on Ascend is now a reported plan, not a fact: on November 3, 2026 Caiwen and Star Market Daily reported, citing industry sources, that V5 will pretrain on Huawei Ascend chips, with a release likely in 2027.
- Beats Astra has no source. The card attached to the viral post is DeepSeek’s logo with v5 typed beside it, with no benchmark in it.
§ 01What the record says, claim by claim
| Claim on the card | What the record says | Grade |
|---|---|---|
| Imminent V5 launch | One Bank of America forecast of the second half of 2026 (36Kr, Sep 2); Caiwen reported on Nov 3 that industry sources expect 2027 | Contradicted by later reporting |
| Liang calls it the biggest bet | He told investors training on Huawei chips is one of the company’s biggest bets (The Information, Sep 21) | About chips, not V5 |
| 2 trillion parameters | A model of about 2 trillion parameters is in training (The Information via 36Kr, Sep 23) | Reported, not called V5 |
| Trained fully on Huawei Ascend | Caiwen and Star Market Daily (Nov 3) report V5 plans to pretrain on Ascend; Caiwen names the Ascend 950 and says training is still in selection | A reported plan, not done |
| About 4x more Ascend chips than Nvidia | Traces to an estimate Liang gave investors for a hypothetical training run, relayed by TrendForce citing Chosun Biz | An estimate, not a V5 fact |
| Keeps open weights | Recent DeepSeek models ship open weights; no V5 statement | Pattern, not a promise |
| Beats GPT-6 Astra | No benchmark, sheet or eval page exists | No source |
§ 02November 3: two reports put V5 on Ascend
Two Chinese financial outlets reported on November 3, 2026 that DeepSeek V5 will be the first DeepSeek model pretrained on Huawei Ascend chips. Both cite people in the chip supply chain. Neither cites DeepSeek or Huawei, and neither company has commented.
| Detail | Caiwen (Zhang Xiaodi) | Star Market Daily, CLS (Huang Xinyi) |
|---|---|---|
| Published | Nov 3, 10:57 Beijing time | Nov 3 |
| Claim | V5 plans to use the Ascend 950 as its core training compute | V5 plans to pretrain on Ascend chips |
| Chip named | Ascend 950 | None |
| Timing | Project points to 2027; release likely in the first or second half, depending on training | None given |
| Stage | Training still in the selection phase | A plan |
| Until now | Ascend ran DeepSeek inference only, not training | Ascend ran DeepSeek inference |
Caiwen adds that DeepSeek and Huawei engineers already adapted low-level operators during V4, that the planned Ulanqab data center is still early and may not host V5’s first training, that DeepSeek currently trains with no other Chinese chip vendor, and that Huawei holds no stake in DeepSeek. Star Market Daily frames the stakes: no top Chinese model has yet been pretrained on domestic chips, and a V4 Pro run on Ascend 910C chips this autumn was post-training only.
What it changes for the card: training on Ascend moves from unsupported to a reported plan for a model that is not trained yet, and the 2027 window contradicts the card’s imminent launch. Both fit the earlier record: Bloomberg’s September order was mainly for running models, and The Information expected Huawei training chips from the fourth quarter. What it does not change: no V5 score, no parameter count and no word from DeepSeek.
§ 03Where each claim came from
The V5 timing. 36Kr’s September 2 roundup of a Bank of America outlook says: “DeepSeek V5 and the upgraded version of ByteDance’s Seed series are expected to debut in the second half of 2026”. That is a bank’s expectation about a quarter, not a report from inside DeepSeek.
The 2 trillion parameters. 36Kr, on September 23: “according to The Information, DeepSeek is currently training a new model with about 2 trillion parameters”. The same piece says Liang Wenfeng told a recent investor meeting that “the company’s next plan is to continue pushing the model to 8 trillion parameters”. For scale, TrendForce, citing Chosun Biz, puts Qwen 3.8 Max at 2.4 trillion parameters and Kimi K3 at 2.8 trillion. A 2 trillion model would not be the largest Chinese model. None of these reports gives the model a name.
The biggest bet. The Information’s own post, September 21 at 15:30 UTC: “training AI models on Huawei chips is one of the company’s biggest bets”. The bet is the chip strategy. The card moved the words onto V5.
Trained fully on Ascend. The same post says DeepSeek expects “a new batch of Huawei training chips as early as the fourth quarter”. A model already trained fully on those chips would need them before they ship. Bloomberg’s September 4 report on DeepSeek’s order for at least 160,000 Ascend chips is summarized by TrendForce this way: “At the time, the chips were expected to be used primarily to run AI models rather than for training.”
Beats Astra. Nothing in the post, the image or any report carries a V5 score. Replies to the post asked for evidence and pointed at the image within hours.
§ 04The chain, dated
§ 05September 30: DeepSeek ships its Ascend software
On September 30, 2026, DeepSeek published Huawei Ascend versions of its core libraries on GitHub. The FlashMLA write-up opens: “On September 30, 2026, DeepSeek open-sourced its inference infrastructure and core components for the Huawei Ascend platform”.
| Library | What it does on Ascend |
|---|---|
| FlashMLA | Sparse attention kernels for the Ascend 950: up to 410 TFlops in prefill (95% of peak) and 360 TFlops in decoding (83%) on DeepSeek V4.1 workloads |
| DeepGEMM-Ascend | Matrix multiplication kernels, API-compatible with DeepGEMM, validated on the Ascend 950 series |
| DeepEP-Ascend | Expert-parallel communication for machine learning training and inference on Ascend NPUs |
| TileKernels | A second backend chosen at runtime, “so the same Python APIs run on both NVIDIA GPUs and Huawei NPUs” |
The TileKernels README thanks “Huawei for its technical support and engineering expertise throughout the development of Tile Kernels’ Ascend backend.” Reuters and Bloomberg reported the release as a DeepSeek and Huawei partnership.
What it changes for the card: the Huawei-chip bet now has code behind it, and DeepSeek V4.1 runs on Ascend 950 with kernels near the hardware’s peak. What it does not change: none of the four releases mentions V5, and none says a model was trained on Ascend. The grades in the table above stand.
§ 06DeepSeek’s own plan points at V4.1 Pro first
DeepSeek’s September 10 launch called V4.1 Flash the smallest model in a new architecture family, and its pricing page names a V4.1 Pro. On DeepSeek’s own stated order, V4.1 Pro is the next release, and a V5 would follow it. That is our reading of DeepSeek’s pages, not something DeepSeek has said about V5. Our V4.1 Flash record has the full launch, the prices and the claims we graded then.
§ 07Where CellCog stands
Our conflict, declared: we build CellCog, a platform where a business hires AI employees that keep memory and work as a team, and we choose the models underneath them. CellCog does not route to DeepSeek today. If V5 ships and earns a place under our employees, we will say so the day it happens, with receipts.
§ 08What we are watching for
- A V5 or V4.1 Pro entry on DeepSeek’s news page or API change log, and a new repository under deepseek-ai on Hugging Face.
- A DeepSeek post on X from its own account.
- A new DeepSeek model id on OpenRouter.
- Any named report that ties the 2 trillion model to the V5 name.
- Confirmation from DeepSeek or Huawei of the Ascend pretraining plan.
§ 09Update log
- September 26, 2026: page opened after the viral V5 card. Seven claims graded; no DeepSeek statement.
- October 1, 2026: added DeepSeek’s September 30 Ascend software release (FlashMLA, DeepGEMM-Ascend, DeepEP-Ascend, TileKernels). No grade changed; still no DeepSeek statement about V5.
- November 3, 2026: added the Caiwen and Star Market Daily reports that V5 plans to pretrain on Huawei Ascend (Caiwen: Ascend 950, 2027). Two grades changed: Ascend training is now a reported plan, and the imminent launch is contradicted. Still no DeepSeek statement.
§ 10Sources
- Caiwen, November 3, 2026 (Zhang Xiaodi): report
-
Star Market Daily (CLS), November 3, 2026 (Huang Xinyi): report
-
The Information on X, September 21, 2026: post
- 36Kr, September 2, 2026 (Bank of America outlook): article
- 36Kr, September 23, 2026 (citing The Information): article
- TrendForce, September 23, 2026 (citing Bloomberg and Chosun Biz): article
- DeepSeek news page: deepseek.com/en/news
- DeepSeek API change log: api-docs.deepseek.com/updates
- The viral post: x.com
Q1Did DeepSeek announce V5?
No. DeepSeek’s news page lists V4.1 Flash as its latest release, dated September 10, 2026, and its API change log has no V5 entry. DeepSeek said at that launch that V4.1 Flash is the smallest model in a new architecture family and that a V4.1 Pro is coming.
Q2Is DeepSeek training a 2 trillion parameter model?
The Information reported it, as relayed by 36Kr on September 23, 2026: DeepSeek is training a new model with about 2 trillion parameters, and Liang Wenfeng told investors the next plan is to push to 8 trillion. DeepSeek has not confirmed it, and no report names that model V5.
Q3Was DeepSeek V5 trained on Huawei Ascend chips?
Not yet. On November 3, 2026, Caiwen and Star Market Daily reported, citing industry-chain sources, that DeepSeek plans to pretrain V5 on Huawei Ascend chips. Caiwen names the Ascend 950, points the project at 2027 and says training is still in the selection phase. DeepSeek and Huawei have not confirmed it.
Q4Does DeepSeek V5 beat GPT-6 Astra?
There is no benchmark to judge. The viral post’s image is the DeepSeek logo with v5 set beside it, and no report, leak sheet or eval page shows a V5 score. Replies to the post asked for evidence within hours.
Q5Will DeepSeek V5 be open weights?
DeepSeek has released open weights for its recent models, including V4.1 Flash under the MIT license. It has said nothing about V5 licensing, so open weights is a pattern, not a promise.