DeepSeek Official Tracker
Source-linked DeepSeek news: Harness desktop and CLI, model releases, API changes, and clearly labelled media, community and rumour coverage.
Official channels and product pages checked every 6 hours. Public signals carry source and confidence labels; AI research passes independent source review via PR. Independent community project, not affiliated with DeepSeek.
English | 中文 · 📡 FEED.md · 📰 Website News · 🤗 HuggingFace · 🔎 Signals · 🤖 Insights · 🌐 Pages
Sister Project: Awesome DeepSeek Harness — curated plugins & ecosystem for dsh (DeepSeek Harness) · Official: platform.deepseek.com | api-docs.deepseek.com | deepseek.com
TL;DR
See the latest DeepSeek news on the website. FEED.md aggregates API changelog, API news, deepseek.com blog, GitHub releases/tags, npm, and Hugging Face models. Product observations track the Harness desktop page separately; SIGNALS.md adds attributed media, community and unverified leads. Collection runs every six hours, subject to upstream availability and GitHub Actions scheduling.
For: people following DeepSeek news and Harness desktop updates · API users tracking pricing and deprecations · researchers watching model weights · developers following CLI releases and the plugin ecosystem.
🔥 Latest — v0.2.1-rc.5 (2026-10-29)
Prerelease / preview · v0.2.1-rc.5
latest
Harness desktop, release tracking and community news
Track DeepSeek Harness desktop (Windows/macOS), dsh CLI releases, DeepSeek models, API changes and model weights. The official product page is monitored directly at deepseek.com/harness, alongside GitHub, npm, HuggingFace and the official news feeds. Product observations distinguish first-seen dates from launch dates.
Media, community discussion and unverified rumours live in SIGNALS.md; reviewed research lives in insights.md. They are never presented as official releases.
📑 Table of Contents
📡 Tracked Sources — 6 official channels
Design principle: Two layers with different guarantees. The official feed (FEED.md) is official-only — no secondary news, no rumours, no LLM rewriting, every line a [Source] you can verify in 1 click. The signals layer (SIGNALS.md) collects community, media and rumour material separately, with each tier stating how far it can be trusted. Nothing crosses between them.
| # | Source | What it catches | Output | Frequency |
|---|
| 1 | DeepSeek API Change Log | Model GA/beta, pricing, deprecation, API feature flags | api-changelog.md | 6h |
| 2 | DeepSeek API News | Long-form official news (news260813 etc) | NEWS.md | 6h |
| 3 | DeepSeek Official Website Blog | Research blog, open-source announcements (V4-Preview, V3.2, Terminus…) — often 1-2 days earlier than API docs | website-news.md | 6h |
| 4 | GitHub deepseek-ai/* Releases & Tags | 33 repos: DeepSeek-V3/R1/OCR/VL2/Coder-V2/Math-V2/Prover-V2, FlashMLA/DeepEP/DeepGEMM/3FS/smallpond, deepseek-harness… | releases.md | 6h |
| 5 | Hugging Face deepseek-ai | New weights & checkpoints (, …), likes/downloads |
🔎 Signals — community, media & rumours
An official-only tracker tells you what already happened. This layer tells you what people think is coming — with the confidence stated up front, so speculation is never mistaken for a release.
SIGNALS.md is regenerated every 6 hours by scripts/signals.mjs over 19 key-free corners with no LLM in that path, so it cannot hallucinate and does not wait on a review to land. Rows are grouped by trust tier:
| Tier | Sources | Can you assert it? |
|---|
rumor | leak wires, 传闻/泄露 aggregates, "next model" speculation | No. Stamped 疑似 / unverified, dated, with what would confirm it |
community | HackerNews, Reddit, X, V2EX, third-party GitHub ports | No — leads, does not prove |
secondary | Established tech media, arXiv, OpenRouter | Authoritative, but not first-party |
A rumour is recorded with its date rather than dropped or quietly promoted. When DeepSeek actually ships, it lands in FEED.md as an official release and supersedes the rumour — the separate tiers make sure the two are never confused. The AI narrative in insights.md layers verified findings on top, reviewed via PR.
Why 6 and not 1? Previously only API docs were watched. Now we cover the full official surface — our audit found deepseek.com blog publishes ahead of API docs (e.g., V3.2), and Hugging Face weights often land hours before a blog post. Single-source = you miss it.
🧭 How it Works
api-docs.deepseek.com/updates ─┐
api-docs.deepseek.com/news ─┤
www.deepseek.com/en/news ─┼─► scripts/track.mjs ─► append-only *.md ─► FEED.md + README Latest (auto)
github.com/deepseek-ai/* ─┤ ▲ │
huggingface.co/deepseek-ai ─┤ │ state.json (seen IDs) └─► git commit & push (retry+rebase)
registry.npmjs.org/@deepseek… ─┘ └─ retry+timeout fetch, date validation, n/a handling
| Stage | Behavior |
|---|
| Official collection | Six release channels plus separate Harness product-page observations; original titles and source links |
| Public signals | Media, community and rumours retain their attribution and uncertainty labels |
| Research | Free-model discovery produces a PR; an independent source review and site checks gate merging |
| Schedule | Collection every six hours, research daily, pending reviews retried hourly; outages can delay updates |
| Dependencies | Node.js collectors, the workflow GitHub token, keyless search services and current free OpenCode models |
🔔 Subscribe / Use
1. Watch research discussions: GitHub → Watch → Custom → Pull requests to follow research drafts and review activity. Notification delivery follows your GitHub settings; watching does not send an email for every feed commit.
2. Bookmark the feed: Single file FEED.md is newest-first, 80 items, with valid YYYY-MM-DD dates (n/a only for GitHub tags without a timestamp). Ideal for quick scan.
3. Consume as data:
# All timelines as markdown
curl -s https://raw.githubusercontent.com/awesome-deepseekharness/deepseek-official-tracker/main/FEED.md
# Raw state for programmatic use (seen IDs)
curl -s https://raw.githubusercontent.com/awesome-deepseekharness/deepseek-official-tracker/main/data/state.json | jq .
4. Feed into your own bot: Re-use scripts/track.mjs — set GITHUB_TOKEN for higher GitHub rate limits, or HF_TOKEN if you mirror HF more aggressively.
⚙️ Automation & Reliability
- Schedule:
cron: 0 */6 * * * (every 6h) + workflow_dispatch + push on scripts/** (so a fix instantly re-crawls).
- Past failures: 2 early
rejected pushes on 2026-08-15 09:1x UTC due to concurrent git push race. Fixed: fetch-depth:0, pull --rebase --autostash with 3-retry loop (scripts/track.mjs:fetchText also retries 3× with backoff, 15s timeout). Since fix: consecutive successes (see Actions tab).
- State:
data/state.json dedupes IDs (live counts — see file). Corrupted entries auto-heal (date validation YYYY-MM-DD, tag vs release dedup).
- README showcase:
🔥 Latest + Previous highlights are auto-synced by scripts/track.mjs:syncReadmes from all official channels and product observations (verbatim titles + excerpts, no LLM) — no manual bump needed.
- No timestamp churn:
FEED.md header is stable except last update: ISO timestamp; no empty commits.
🛠️ Manual Run
git clone https://github.com/awesome-deepseekharness/deepseek-official-tracker
cd deepseek-official-tracker
node scripts/track.mjs # needs Node 22+; no deps
# optional: GITHUB_TOKEN=ghp_xxx node scripts/track.mjs # lifts GitHub 60 req/h limit
Files are append-only; re-running without new upstream items is a no-op (README Latest sync is idempotent).
❓ FAQ
Q: Is this affiliated with DeepSeek? No. Data belongs to their respective owners (DeepSeek, GitHub, npm, Hugging Face). We only mirror with citations under CC0-1.0.
Q: Why is npm @deepseek-ai/dsh tracked but not PyPI? DeepSeek's official dsh distributes via npm today. PyPI (deepseek-ai org) can be added on request — open an issue.
Q: Do you rewrite titles with AI? No. We keep official titles verbatim (entity-decoded) + 600-char summary from changelog HTML, with [Source] link. Zero LLM in the pipeline.
Q: GitHub tags show n/a date? GitHub tags API has no timestamp (only releases do). We keep n/a and sort them last in FEED.md — avoids fabricating 2012-26- style bugs (fixed in 2026-08-27).
Q: How do I get notified instantly? Use GitHub Watch → Custom → Pull requests and Releases still not instant? Better: enable Watch on Discussions or use an RSS bridge: https://github.com/awesome-deepseekharness/deepseek-official-tracker/commits/main.atom.
Q: Can I add more deepseek-ai repos? Yes, edit scripts/track.mjs:OFFICIAL_REPOS (33 today) and PR. We track active repos; archived forks are excluded.
Q: HuggingFace shows likes and downloads — is it live? Yes, pulled from huggingface.co/api/models?author=deepseek-ai. Sorted by lastModified.
🔍 Keywords / SEO
deepseek deepseek api deepseek oficial tracker deepseek news deepseek changelog deepseek v4 deepseek v4.1 deepseek v4.1 flash deepseek v4 pro deepseek v4 flash deepseek v3.2 deepseek v3.2 exp deepseek r1 deepseek r1 0528 deepseek v3 deepseek harness dsh deepseek deepseek github releases deepseek huggingface deepseek npm deepseek pricing deepseek context caching deepseek 1m context deepseek api docs
🤝 Contributing & Related
- PRs welcome: fixing a slug, adding a repo, improving
fetchText resiliency. The script is vanilla Node.js (no deps) for easy review.
- Issues: Report a missed official channel — we aim to stay within 6h of any
deepseek-ai or deepseek.com publication.
- Related:
📄 License
CC0-1.0 — Data belongs to their respective sources; this repo only mirrors and organizes them automatically. No affiliation with DeepSeek.
