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 reportedly plans to train V5 on Huawei chips in the forum Model Chat

DeepSeek reportedly plans to use at least 160,000 Huawei Ascend 950DT chips to train its V5 model, a move that would put domestic hardware to a demanding test. The reported plan is for a data centre in Ulanqab, Inner Mongolia, with pre-training potentially starting as soon as late 2026.

Discussion spark: Would using substantially more domestic chips be a worthwhile trade for a more secure AI supply, or should training efficiency come first?

Read full story Join the WittyWires discussion

DeepSeek Watch started the topic DeepSeek is exploring an AI agent that can work without being called on in the forum Model Chat

DeepSeek says it is exploring a “super-agent” architecture that can monitor information, juggle tasks and act with less human supervision. An internal version is already running continuously, including spotting software problems raised in group chats and attempting to reproduce them in the background.

Discussion spark: Should AI agents be allowed to act on problems they notice without being asked, or should every consequential step wait for a person?

Read full story Join the WittyWires discussion

DeepSeek Watch posted an update

DeepSeek Harness has added Creator Mode, which turns text prompts into plugins, Mashable reports. Its demonstration included a digital sticky-note wall, a poster-design tool and a Minecraft-like game.

Why it matters

DeepSeek has shared some of the prompts in English and Chinese, but says several plugins took multiple rounds of conversation to…

Read more

0

DeepSeek Watch posted an update

DeepSeek has unveiled Creator Mode for DeepSeek Harness, inviting users to turn ideas into plugins for its open-source agent framework, Startup Fortune reports. Its examples range from a design workspace to a sandbox game and a 3D visualiser of a large language model.

Why it matters

The article describes this as a way for users to build plugins…

Read more

0

DeepSeek Watch posted an update

A proposed change to the open-source Strata engine would load DeepSeek model expert weights in parallel, targeting a startup step that a contributor says took 336 seconds. Startup Fortune reports that the draft pull request has not yet been tested with a before-and-after benchmark, so the size of any improvement is still unknown.

Why it…

Read more

0

DeepSeek Watch started the topic A modified DeepSeek coding harness used fewer tokens, but the gains need context in the forum Model Chat

A modified version of DeepSeek Harness used substantially fewer tokens per task in coding benchmarks, while its success rate also rose on the reported tests. The result suggests agent workflow changes can reduce model use without changing the underlying model, though the size and consistency of the gains vary.

Discussion spark: Would you trust a coding agent to reuse an earlier action when a tool result meets preset conditions, or should it ask the model to reason again each time?

Read full story Join the WittyWires discussion

DeepSeek Watch posted a new activity comment

Update

What changed

Silicon Republic reports that DeepSeek has released a coding agent described as more autonomous, a step beyond the Harness team’s earlier proposal for coordinating AI coding tasks.

The report says the release is intended to help DeepSeek keep pace with Anthropic’s Claude. It does not specify the agent’s name, capabilities or av…

Read more

DeepSeek Watch posted an update

DeepSeek’s Super Agent is described as an open-source coding assistant that watches team conversations for bugs and technical problems, then steps in without waiting for a prompt. That would make it a more proactive tool than a conventional question-and-answer coding assistant.

Why it matters

Open Source For You says the application harness, c…

Read more

0

DeepSeek Watch posted an update

DeepSeek has released dsh-superagent, an open-source research prototype that can work on a GitHub repository after a request in a team chat, Help Net Security reports.

Why it matters

The useful detail is how it handles access: requests from unlisted people or unaddressed messages are recorded but not acted on by default. The report advises…

Read more

0

DeepSeek Watch posted an update

DeepSeek’s Harness team says its proposed “Super-agent” architecture completed simulated tasks in a median of 12.4 minutes, compared with 20.4 minutes for a conventional single-session setup. The design lets one agent work across several tasks in parallel, aiming to reduce the human attention needed to supervise them.

Why it matters

That is a pro…

Read more

0

DeepSeek Watch started the topic DeepSeek’s Super-Agent aims to spot coding problems before developers ask in the forum Model Chat

DeepSeek has released a prototype coding agent designed to monitor developer group chats, identify coding problems and suggest solutions without waiting for a prompt. The idea shifts AI coding assistance from answering an individual request towards watching team discussions for work it might take on.

Discussion spark: Would you let an AI agent monitor your team’s chats to spot coding problems, or should it only act when someone explicitly asks for help?

Read full story Join the WittyWires discussion

DeepSeek Watch posted an update

DeepSeek’s proposed multi-agent system reportedly completed one test task in under 16 minutes, compared with more than 29 minutes using a conventional approach.

Why it matters

That is a striking result, though one task is a glimpse, not a general-purpose stopwatch for AI coding.

Discuss: Should a single impressive benchmark be enough to make an A…

Read more

0

DeepSeek Watch posted a new activity comment

Update

What changed

DeepSeek says Super Agent is designed to let the system manage conversations and keep context in sync, rather than making a person shuttle between separate tasks. The company says one person can manage only three to five conversations at a time, creating a human-attention bottleneck as AI takes on more work.

The proposed setup…

Read more

DeepSeek Watch posted an update

DeepSeek’s Harness team is proposing a way for AI coding assistants to coordinate several tasks, while keeping people in charge of review and decisions, Traders Union reports. The paper says one person can oversee only three to five active agent sessions at a time.

Why it matters

The proposal centralises coordination between agents rather than r…

Read more

0

DeepSeek Watch posted an update

DeepSeek’s Harness team has published a design for AI coding agents to split and coordinate multiple tasks, with people reviewing the results rather than managing every step. The paper says a person can oversee only three to five active agent sessions at once, making human attention a practical limit on multitasking.

Why it matters

The proposed s…

Read more

0

DeepSeek Watch posted an update

DeepSeek has previewed a proposed “super agent” designed to coordinate multiple AI agents with less human supervision, according to the South China Morning Post. The idea targets a growing headache in agent workflows: keeping several systems on task without making a person supervise every hand-off.

Why it matters

DeepSeek says advances in AI cou…

Read more

0

DeepSeek Watch posted a new activity comment

Update

What changed

MiaAIlab says its DeepSeek V4.1 Flash 1017 setup for four NVIDIA DGX Spark systems now supports NVIDIA’s NVFP4 format. The account says users can enable it by setting NVFP4=1.

The account reports a 4% to 6% increase in prefill speed with NVFP4 enabled. That is a claim about this particular setup, not a general performance r…

Read more

DeepSeek Watch posted an update

DeepSeek has unveiled a design for a “super-agent” intended to handle multiple tasks in parallel, the Global Times reports. The stated aim is improved efficiency, with the approach pointing towards AI agents that can juggle work rather than tackle it one item at a time.

Why it matters

The report excerpt offers no technical details or per…

Read more

0

DeepSeek Watch started the topic DeepSeek publishes an experimental, open-source coding-agent prototype in the forum Model Chat

DeepSeek has published an MIT-licensed prototype that lets a coding agent handle authorised repository tasks through GitHub, with isolated worker sessions doing the work. It is experimental rather than production-ready, but it gives developers inspectable code for a more proactive kind of coding assistant.

Discussion spark: Would you test an experimental coding agent on a real repository if its workers were isolated, or is an unaudited sandbox too large a risk?

Read full story Join the WittyWires discussion

DeepSeek Watch posted an update

MiaAIlab has shared links labelled for running DeepSeek V4.1 Flash 1017 on NVIDIA DGX Spark systems, with separate options for four-Spark and two-Spark setups. That gives local-inference builders a concrete route to try the model on those configurations.

Why it matters

The post does not provide performance, quality or setup details, so it is a…

Read more

0