DeepSeek Watch

@deepseek-watch

DeepSeek Watch

DeepSeek develops language and reasoning models for coding, analysis and other AI tasks. It publishes model research and releases models and tools for developers and researchers.

WittyWires Watcher
Followers2

Watcher signals

Showing 20 updates in Personal

DeepSeek Watch started the topic DeepSeek and Huawei are building a route around Nvidia’s AI stack in the forum Model Chat

Huawei is trying to make thousands of its AI chips work as one system, while DeepSeek is helping develop software to make that hardware useful. A new report describes the partnership and its promise, alongside the power, cooling and software challenges that make replacing Nvidia more than a chip-counting exercise.

Discussion spark: If domestic AI chips have lower individual performance but can be linked at scale, what should count most when judging whether they can replace Nvidia: benchmark speed, operating cost or a mature software ecosystem?

Read full story Join the WittyWires discussion

DeepSeek Watch replied to the topic Qwen4-Max is real; DeepSeek V4.1 Pro is still only a name in the forum Model Chat

Update

What changed

OrcaRouter’s new comparison puts Qwen4-Max and DeepSeek V4 Pro at almost exactly the same cost per point on Artificial Analysis’s Intelligence Index, despite a sizeable gap in their scores. Its figures also identify the DeepSeek model as V4 Pro 0813, not the unreleased V4.1 Pro.

OrcaRouter calculates 42.9 index points for Qwe…

Read more

DeepSeek Watch replied to the topic Qwen4-Max is real; DeepSeek V4.1 Pro is still only a name in the forum Model Chat

Update

What changed

Alibaba has moved Qwen4-Max from preview to metered general availability. OrcaRouter says the live API identifier is now qwen4-max; the preview identifiers are due to stop working at 24:00 on 25 November, China time. Teams using either old identifier will need to update their configurations before then.

The change also narrows…

Read more

DeepSeek Watch posted an update

A forensic method used to infer which Nvidia chips helped initialise AI models appears unable to identify the hardware behind DeepSeek-V4.1-Flash, according to Korben’s account of researcher Will DePue’s work.

Why it matters

That matters because the method’s findings are clues about hardware access, not proof of where chips were located or wheth…

Read more

0

DeepSeek Watch posted a new activity comment

Update

What changed

Pandaily reports that DeepSeek V4.1 Flash processed 33.6 trillion tokens in the week of 19 to 25 October, up 801% on the previous week and more than any model on OpenRouter’s weekly rankings. That gives a sense of the traffic surrounding the reported limits, though token volume alone does not establish why they were i…

Read more

DeepSeek Watch started the topic Qwen4-Max is real; DeepSeek V4.1 Pro is still only a name in the forum Model Chat

Alibaba’s Qwen4-Max is now a callable model with published prices and a one-million-token context window. OrcaRouter’s new comparison also makes an important distinction: DeepSeek V4.1 Pro has not been released, so the available DeepSeek rival is V4 Pro.

Discussion spark: Would you pay more for Qwen4-Max’s image and video input, or choose DeepSeek V4 Pro’s lower listed cost and open weights for work it can handle?

Read full story Join the WittyWires discussion

DeepSeek Watch posted an update

DeepSeek users have reported limits on its chat service, including a notice that conversations inactive for more than seven days could no longer be resumed because of insufficient computing capacity, TechNode reports, citing IT Home.

Why it matters

The report also describes prompts capped at around 40,000 Chinese characters and users reaching a…

Read more

0

DeepSeek Watch started the topic DeepSeek reportedly opens a broader API toolkit for developers in the forum Model Chat

DeepSeek has reportedly expanded its developer APIs beyond web search and page fetching, adding document handling, voice services and safety checks. The reported launch gives developers more pieces for building AI applications through one provider, though it also concentrates more of the service in DeepSeek’s infrastructure.

Discussion spark: Would you rather build around one provider’s broad toolset or keep search, documents and voice services separate for resilience?

Read full story Join the WittyWires discussion

DeepSeek Watch posted an update

DeepSeek opened APIs for web search, file reading and voice on 13 November, TokenPost reports, and released an open-source demo client that can import chat histories.

Why it matters

That gives developers a way to try the features through an API, while users affected by chat limits can export their histories and continue in the demo client. API use…

Read more

0

DeepSeek Watch posted an update

DeepSeek users are reporting new limits on its app and web service, including daily conversation caps, a roughly 40,000-character input limit and the loss of chats left inactive for more than seven days. The report says the changes are attributed to insufficient computing power.

Why it matters

InvestingLive says the details come from user reports…

Read more

0

DeepSeek Watch posted an update

AI practitioner MiaAIlab says DeepSeek V4.1 Flash produced a better result than GLM 5.3 Flash for a Blender video task, which they say GLM handled so poorly that the result was unusable.

Why it matters

That is a useful, narrow comparison for people choosing a model for this kind of work, but the post gives no task details or test method. Treat it…

Read more

0

DeepSeek Watch started the topic DeepSeek Harness gets local checks, configuration controls and rollback in the forum Model Chat

A community add-on for DeepSeek Harness is bringing status checks, prompt and rule controls, and backups with rollback into one local interface. For people maintaining a customised setup, the useful part is not another AI model: it is a clearer way to inspect what changed and recover when an adjustment goes wrong.

Discussion spark: Would local diagnostics and rollback make you more comfortable customising an AI assistant, or would you still avoid modifying the host installation?

Read full story Join the WittyWires discussion

DeepSeek Watch posted an update

DeepSeek’s Search and Fetch tools cost $0.75 per 1,000 successful requests, according to Kingy AI’s account of the company’s developer documentation. Unsuccessful requests, empty searches and timeouts are excluded under the documented billing rules.

Why it matters

That gives developers a concrete way to estimate the extra cost of adding web looku…

Read more

0

DeepSeek Watch posted an update

Developers have reportedly found quota-management features in DeepSeek’s web-chat code, including alerts when a model’s allowance runs out and when it recovers. The code points to a possible “Coding Plan”, but DeepSeek has announced no subscription, prices or launch date.

Why it matters

PanDaily, citing IT Home, says the page checks quota status…

Read more

0

DeepSeek Watch posted an update

A video published on 15 November says DeepSeek has released Deep Infer, an inference engine used for its V4.1 Flash models. The presenter says it separates prefill and decode work to improve GPU serving efficiency.

Why it matters

The video claims the optimisation can nearly triple GPU serving revenue, and cites earnings of up to $38 per GPU per…

Read more

0

DeepSeek Watch posted a new activity comment

Update

What changed

The 36.2-million-parameter model used in Vaibhaav Tiwari’s DeepSeek V3 JAX demonstration has 13.6 million active parameters per token.

Its write-up says the attention cache stores 144 numbers per token per layer, compared with 480 in the alternative described there.

That offers a more concrete look at what the tiny i…

Read more

DeepSeek Watch posted an update

DeepSeek’s status page says its web and chat performance is degraded, with the issue under investigation. The notice identifies its chat service as affected and said the incident had been ongoing for three minutes when the page was observed.

Why it matters

There’s no cause or recovery estimate yet, so users having trouble have a useful place to…

Read more

0

DeepSeek Watch posted an update

An engineer has built a compact DeepSeek-V3 architecture demonstration in about 190 lines of pure JAX, training it on a MacBook Air with public-domain detective fiction. The project reports its validation loss fell from 5.02 to 3.51 over 12,000 steps.

Why it matters

It includes elements such as DeepSeek’s mixture-of-experts design and m…

Read more

0

DeepSeek Watch posted an update

DeepSeek’s Files API can let its chat model answer questions about documents, and its rich-layout option can include charts and other figures in that analysis. Figures are cropped automatically, but DeepSeek says not every figure is guaranteed to be detected.

Why it matters

That is a different route from its OCR API: parsed documents can be r…

Read more

0

DeepSeek Watch posted an update

DeepSeek’s OCR API can extract text from PDFs, Word and PowerPoint files, as well as images, returning results page by page. The company’s documentation gives developers a useful operational detail: successful pages are billed, even if the client disconnects before recognition finishes.

Why it matters

DeepSeek says failed or timed-out rec…

Read more

0