DeepSeek is rationing its free web and app chat. On November 16, 2026, Lianhe Zaobao, InvestingLive and Interface News via Sina Finance describe three simultaneous changes: daily conversation quotas, a ~40,000-character input ceiling, and removal of long threads idle for seven days, all attributed to insufficient compute in user-facing notices.
Is DeepSeek down? No — DeepSeek’s status page still reports normal operations. Is DeepSeek still free? Yes, but “free and unlimited” is no longer the working assumption for heavy chat users.
For builders, the split matters: consumer quotas hit discovery and casual coding, while the metered API track you could wire into agents yesterday is on different pools and pricing. Plan capacity before you bake a free chat tab into a product demo.
What changed on November 16, 2026
Chinese financial press and English-language wires converge on the same three knobs. None of the numeric quotas came from a DeepSeek press release at publish time; treat counts as user-reported unless DeepSeek posts them.
| Change | What users report | Stated reason |
|---|---|---|
| Daily conversation cap | New hard stop after a day’s usage | Server congestion / quota exhausted |
| Single-input length | About 40,000 Chinese characters (~roughly 40k characters in Latin text depending on encoding) | Compute per request |
| Idle long chats | Threads inactive seven days (some users cite 72 hours) no longer resume | Compute / storage reclaim |
Lianhe Zaobao quotes the quota message in English: “Usage limit reached due to high server demand.” That wording is sharper than the generic “Server busy” banners users saw during October and November incidents — it names a quota, not just backoff.
An hourglass of tokens evoking DeepSeek daily conversation quotas on free chat
Interface News, syndicated on Sina Finance, adds engineering context: idle long conversations hoard KV cache in GPU memory, so purging week-old threads is partly HBM recovery to serve active users. That is plausible for any long-context chat stack, but DeepSeek has not published a technical blog to confirm the implementation detail.
Daily token meter evoking DeepSeek usage limits on free chat
Official status versus user experience
At check time on November 16, 2026, status.deepseek.com listed:
- Chat service at 99.56% uptime from August through November 2026
- V4.1 Flash API at 99.56%
- V4 Pro API at 99.74%
- A green banner: “Everything is running smoothly”
So the limits are policy and scheduling, not the red incident calendar entries from October 15 (a 4h52m degradation affecting chat and APIs) or November 7 partial outages reporters flagged to Aju Press.
The tension is normal for hyperscale chat: availability metrics stay green while per-user fairness rules tighten. If you run on-call for an app that scrapes the free UI, monitor error text, not just HTTP 200.
Why demand may have outrun supply
Interface News and 10jqka field reporting sketch a demand curve:
- OpenRouter weekly tokens for DeepSeek V4.1 Flash reportedly jumped to 33.6 trillion in late October, with ~801% week-on-week growth in one cited window
- Eight logged incidents since October on DeepSeek’s status timeline
- A full application API launch on November 13, 2026, letting developers mirror App features — search, files, voice — on paid meters
Meanwhile supply-side stories point to long-lead data-center builds. Interface News cites Tech Daily reporting on a gigawatt-class Ulanqab site with 160,000 domestic AI chips, with 2027–2028 production — useful context, not relief this week.
Export friction sits in the background. explainx.ai has tracked H200 approval bottlenecks and neocloud access rules. DeepSeek has not tied November 16 quotas to a policy speech; the link is analyst inference, not a primary quote.
Free chat versus API: what builders should run where
| Surface | Billing model (Nov 2026) | Limit style | Best for |
|---|---|---|---|
| Web / App chat | Free C-end | Daily quota + input cap + idle purge | Exploration, light Q&A |
| deepseek-flash / V4 Pro API | Per-token + tools | Concurrency caps; peak/off-peak pricing | Agents, integrations |
| Open weights (when released) | Your GPUs | Your ops budget | Air-gapped or compliance-heavy |
DeepSeek’s peak and off-peak API pricing was already a demand-shaping tool. November 16 adds demand-shaping on the free tier without announcing a consumer subscription to buy through the wall — unlike western assistants that monetize queue-skipping.
A scale weighing free chat limits against metered API token spend for production agents
For agent builders, three practical moves:
- Move production off the free UI — you cannot rely on quota you cannot read from an API header.
- Chunk long documents below the reported 40k input ceiling; treat the free tier as short-turn coding, not repository ingestion.
- Cache exports of threads you still need; seven-day idle deletion breaks “leave a spec in chat and return next sprint” workflows.
Our DeepSeek Harness DAU post shows how fast consumer distribution grew; quotas are the predictable second act when inference is the margin.
Timeline: DeepSeek capacity moves before November 16
Understanding the sequence helps you guess whether quotas are temporary congestion or a new steady state:
| When | Move | Source |
|---|---|---|
| Feb 2025 | Pause on API balance top-ups | Prior DeepSeek notices, cited in field reporting |
| May 2026 | Quiet limits on regenerate / edit | User reports, InvestingLive |
| Aug 2026 onward | Peak / off-peak API pricing | DeepSeek pricing pages |
| Oct 2026 | Repeated chat/API degradations | status.deepseek.com incident log |
| Nov 7, 2026 | Partial chat/API congestion | Aju Press, user reports |
| Nov 13, 2026 | Full application API suite | DeepSeek changelog, our API post |
| Nov 16, 2026 | Daily free-chat quota + idle purge | Zaobao, Interface News, InvestingLive |
None of these rows include a DeepSeek blog post titled “why we capped free chat.” Until that lands, attribute limits to what users see and what financial media verified with multiple testers.
How this compares to other Chinese flash models
DeepSeek built its brand on aggressive API pricing and a no-paid-consumer-tier posture. Tencent’s Hy4 flash launch the same day pushes WorkBuddy free promos and half-price TokenHub APIs — a different subsidy strategy, still compute-bound at scale.
On benchmarks, Vals CyberOffense and our frontier comparison table show V4.1 Flash as the model developers route to when cost per task matters. Quotas on the free UI do not change API leaderboard scores — they change where students and hobbyists first touch the model.
If you need local inference, see Baseten’s ultrafast DeepSeek engine and GGUF stack guides for hardware you control.
What to watch next
DeepSeek has not said when quotas reset (midnight Beijing, rolling 24 hours, or per-account calendars are all unanswered), whether WorkBuddy-style partners get exempt pools, or if a paid consumer tier is coming.
Official places to check:
- status.deepseek.com for incident versus policy changes
- DeepSeek pricing and docs for API concurrency and peak/off-peak tables
- DeepSeek’s X or Weibo accounts for a formal quota table — none was linked in the articles we read on November 16
QuestMobile figures cited by Aju Press put DeepSeek at roughly 123 million monthly active users in China as of September 2026. Even a modest per-user daily cap scales to a large inference bill. Treat free chat as marketing spend with a ceiling, and APIs as the product.
Is DeepSeek better than ChatGPT for free coding now?
For heavy daily coding on zero budget, probably not on November 16, 2026 — if the quota reports are accurate. ChatGPT, Gemini, and domestic rivals still offer different limit shapes (slowdowns, model switches, paid Plus tiers). DeepSeek’s shift names a hard daily wall on the free UI, which is new in public reporting.
That does not mean DeepSeek lost coding quality overnight. V4.1 Flash 1017 still ranks strongly on many agent tasks; only the unmetered consumer funnel tightened.
A pattern that still works:
- Plan in a frontier model you pay for or self-host.
- Implement with DeepSeek API or local weights where licenses allow.
- Stop using the free web UI as a CI substitute.
HN-style workflows that ran hundreds of free calls per day were always fragile; November 16 makes the fragility explicit.
Government and enterprise users
If your customer base includes US federal civilian agencies, remember DHS orders to remove DeepSeek — unrelated to quotas but decisive for procurement. Consumer limits do not change that ban; they simply reduce shadow IT usage inside those networks.
For everyone else, re-read data residency terms before pointing regulated documents at either free chat or APIs. Quotas do not reduce compliance obligations.
Workarounds that respect ToS
explainx.ai does not advocate multi-account evasion or bot farms. Legitimate mitigations:
| Approach | When it fits |
|---|---|
| API metering with spend caps | Production agents with billing alerts |
| Off-peak API calls | Batch jobs that can wait for discounted UTC windows |
| Open weights on your GPUs | Air-gap or strict data control |
| Smaller prompts | Stay under reported 40k input limits on free tier |
| Export chats weekly | Avoid seven-day idle loss |
The static browser demo client and Harness marketplace ecosystem remain relevant — they assume API keys, not scraped free sessions.
KV cache and why idle threads die
Interface News’s KV cache explanation matches how transformer serving works: long threads reserve GPU RAM for attention state. Letting a million-token hobby thread idle is expensive when 123M MAU compete for the same pool.
Purging seven-day idle threads is a brutal UX choice but a rational ops choice. Western products usually archive chats in cheap object storage; DeepSeek’s reported approach deletes resumability to reclaim HBM fast.
If your workflow is “keep spec in chat history”, migrate to files in repo + short chat turns — the same lesson as context window cost guides teach for paid APIs.
Incident history versus quota policy
DeepSeek logged eight degradation events from October into November, per Interface News’s reading of the status calendar. The November 16 move is not another red calendar entry — it is product policy.
Teams should tag alerts separately:
- Incident — error rate, latency SLO breach, status page yellow/red
- Quota — HTTP 200 with quota exhausted copy
Mixing them in runbooks causes wrong escalations (page infra when product management already shipped a cap).
Competitive ripple on the same news day
Tencent launched Hy4 flash with free WorkBuddy promos while DeepSeek tightened free chat. Chinese model competition is now bifurcated:
- Subsidize flash APIs to win office automation
- Ration free consumer inference when OpenRouter traffic explodes
Watch OpenRouter share in our Asia token analysis for early signals of routing changes after quota day.
