DSH 0.2 and DeepSeek V4.1 Pro: A Rumoured Double Release, Tested Against DeepSeek's Own Code

On Monday morning in Beijing, the Deep​Seek watcher @teortaxesTex posted that it "would be cool if Deep​Seek released DSH 0.2 together with V4.1 Pro that's further trained for DSH and agent teams on Monday," and then hedged inside the same breath: "I think at least next week until Thursday is very likely." That is a wish with a calendar attached, not a report. At the centre of it sits DeepSeek V4.1 Pro, which still has no model card, no weight file, no price and no endpoint — and the other half of the wish, an 0.2 line of Deep​Seek Harness, has never been published in any channel. The two Deep​Seek models anyone can actually call today are DeepSeek V4.1 Flash, the 552B open-weights release from 10 September, and DeepSeek V4 Pro, the 1.6T flagship that Deep​Seek's changelog says continues past 14 September with its billing unchanged. Neither is what the tweet is about.

So the useful thing to do with a sentence like this is not to argue about the date. It is to separate the parts that can be checked from the part that has been guessed — and this particular wish bundles three separate assertions, two of which are falsifiable in about ten minutes each, and the third of which is genuinely interesting because it points at real work DeepSeek has already published. Here is each of them, in the order they can be tested.

What the signal is, in evidence terms

The tweet is one paragraph from an account that follows DeepSeek closely and has been early before. It was posted at 02:36 UTC on Monday 28 September 2026, which is 10:36 in Beijing — the same morning it predicts a launch for. There is no DeepSeek document behind it, no second source, no developer preview link, and no cost to the author if it is wrong. The account is not describing something it has seen; it is describing a shape it thinks DeepSeek's schedule has.

Three assertions are stacked in it, and they fail or survive independently:

• A version number. That the harness's next release is numbered 0.2 — a minor bump rather than the 0.1.x release candidates DeepSeek has been shipping daily for six weeks.

• A model. That DeepSeek V4.1 Pro ships as a product at all, and that a reader will be able to call it.

• A joint arrival. That the harness bump and the model land together, on a Monday, because the model is "further trained for DSH and agent teams."

The first is checkable against a package registry. The second is checkable against DeepSeek's own changelog, price sheet and Hugging Face organisation. The third is the only one that requires any judgement, and it is also the only one where the tweet's author is pointing at something real.

There is no 0.2 in DeepSeek Harness

DeepSeek Harness is open source under MIT, published on GitHub and npm, and both channels are readable without an account — which makes this the cheapest part of the claim to test.

• npm — @deepseek-ai/dsh has 27 published versions. The highest is 0.1.7-rc.2. There is no 0.2.x anywhere in the registry, and the dist-tags read latest: 0.1.7-rc.2, next: 0.1.7-rc.2, alpha: 0.1.7-alpha.2.

• GitHub — 22 published releases, every one of them marked prerelease. The newest is dsh-v0.1.7-rc.2, published 24 September 2026 at 14:10 UTC. The repository's own root manifest carries the same version.

• The desktop build — DeepSeek ships a packaged Windows and Apple-silicon client through its own download host, and its two production update feeds both report version 0.1.7-rc.2 with a release date of 24 September.

• The cadence — the 0.1 line has moved through alpha and release-candidate builds roughly every one to three days since mid-August: 0.1.6-alpha.1 on 15 September, 0.1.6-alpha.2 on the 17th, 0.1.7-alpha.1 and alpha.2 on the 22nd, rc.1 on the 23rd, rc.2 on the 24th. Nothing about that cadence is building toward a 0.2 tag; it is building toward 0.1.7 stable.

There is one change here a reader can act on, and it is small but real: npm's latest tag now resolves to 0.1.7-rc.2, so npx @deepseek-ai/dsh web installs a newer release candidate than it did a few days ago, when the default still pointed at 0.1.5-rc.3 while the desktop app was already on 0.1.7-rc.2. The desktop build and the npm default have finally converged. That is a packaging event — useful if you are installing this week, irrelevant as evidence about a model.

There is no V4.1 Pro anywhere DeepSeek writes things down

This is the more consequential half, because a rumour about a harness version costs nobody anything, while a rumour about a flagship tier is the kind of thing a team plans a quarter around. DeepSeek maintains four places where a released model appears, and the name is absent from all four.

• The changelog — DeepSeek's own API documentation lists 21 dated entries, from DeepSeek V2.5 in September 2024 to the present. The newest is 10 September 2026, the DeepSeek-V4.1-Flash release. Nothing dated after it, and "V4.1 Pro" does not appear in any entry.

• The price sheet — DeepSeek's pricing page exposes exactly two model names today: deepseek-flash (version DeepSeek-V4.1-Flash) and deepseek-v4-pro (version DeepSeek-V4-Pro-0813), at concurrency limits of 2,500 and 500 respectively. A successor tier that a reader could call would need a row here, and there is no row.

• Hugging Face — the deepseek-ai organisation's newest repository is DeepSeek-V4.1-Flash, created 10 September 2026 at 02:17 UTC, past 650,000 downloads. The entries behind it are from 31 August and 13 August. Nothing newer, and nothing with Pro in the name for the 4.1 generation.

• Serving — on our own side, the OrcaRouter model API returns 404 for DeepSeek V4.1 Pro under both the hyphenated and the dotted spelling, while DeepSeek V4.1 Flash and DeepSeek V4 Pro both resolve. Artificial Analysis has no page for it either. Those are independent of DeepSeek's own documentation and they say the same thing.

The name is not brand new — it surfaced on the vendor side in early September, in reporting around a plan to reroute DeepSeek V4 Pro traffic to V4.1 Flash, and that reporting is where most readers first met it. What DeepSeek's documentation confirms today is narrower and less exciting: the V4 Pro service continues past 14 September, and the retired Flash identifiers are still accepted and are served by DeepSeek V4.1 Flash at Flash pricing. A plan is not an artefact. The file that would settle this is a model card, and there is not one.

The sharpest version of that check sits inside DeepSeek Harness itself. The harness talks to DeepSeek's API through its own adapter, and that adapter ships a default model catalogue — the short list of models the client offers before a user configures anything. It has exactly two entries: deepseek-flash, displayed as DeepSeek-V41-Flash with text and image input, and deepseek-v4-pro, displayed as DeepSeek-V4-Pro. There is no third entry. If DeepSeek were preparing a Pro tier that its own harness was trained or tuned around, this file — two entries long, in the vendor's own public repository — is the most natural place for it to appear first, and it is not there.

The part that is checkable: "further trained for agent teams"

This is the clause worth taking seriously, because it is the only one that maps onto work DeepSeek has actually published, and because it explains why anyone expects a co-release at all.

Start with the paper. On 19 September, DeepSeek submitted DeepSeek Elastic Compute (DSec): A Sandbox Infrastructure for Effective Agentic Training at Scale to arXiv (2609.22978). It describes the platform behind DeepSeek's agent training and evaluation: four sandbox backends — FnCall, container, microVM and full VM — behind one SDK, environment images streamed on demand from the company's distributed filesystem, and lifecycle management co-designed with the reinforcement-learning loop so that stateful rollout execution survives preemptible GPU training. One production unit is around 160 nodes serving roughly three million sandboxes a day, sustaining over 380,000 concurrent sandboxes and more than 5,000 creations per second. One section is devoted to agents that learned to take shortcuts — forging internal requests, reading logs they should not have, and tripping a kernel bug by recursing into /proc — and to the AppArmor and eBPF mitigations that followed.

That is agent-training infrastructure, published in the open, three days before the tweet's window. It is evidence that DeepSeek is building the machinery a "trained for agent teams" claim would need. It is not evidence about a schedule.

Then the harness side. Agent Teams exists in DeepSeek Harness as four plugins published under experimental package names, first pushed to npm on 9 September and now tracking the 0.1.7 line. An architecture note dated 18 September collapsed the previous two-switch arrangement into a single optional bundle that carries the team service, the tools and the browser panel together — and left it disabled by default in the plugin page. The 0.1.7-rc.1 release notes, from 23 September, describe what the panel does now: it shows team members and their tasks live, can be opened or switched from the session page header, and the team tools address members by name. The same notes record a change "adjusting image scaling and token estimation to accommodate DeepSeek V4.1" — the harness tuning itself for the current model, not for a future one.

Finally, the model card. DeepSeek's own card for DeepSeek V4.1 Flash states that its code-agent benchmarks — Terminal-Bench, DeepSWE, NL2Repo-Bench, ProgramBench — were run "with the Minimal mode of DeepSeek Harness and a 1M-token context window." That is a vendor telling you its agent numbers were produced inside its own harness. It also means those numbers are vendor-reported and unreproduced: DS-V4.1-Flash scores 90.6 on Terminal-Bench 2.1 and 74.2 on DeepSWE v1.1 on DeepSeek's own setup, and nobody outside DeepSeek has re-run them. The independent view is thinner and comes from a different date: the Artificial Analysis figures carried on OrcaRouter's model pages put DeepSeek V4.1 Flash at an Intelligence Index of 39.5 (evaluated 10 September) and DeepSeek V4 Pro at an AA Coding Index of 68.8 (evaluated 13 August). Those two rows are from different evaluations days apart in different months, so they are not a like-for-like comparison — read them as two snapshots, not one scoreboard.

Why "on Monday" would be unusual — and why the hedge matters more

DeepSeek's release history is public, and read as a calendar it argues gently against the first half of the tweet.

• Recent releases land Thursday or Friday — 10 September 2026 was a Thursday, 21 August a Friday, 13 August a Thursday, 31 July a Friday, 24 April a Friday.

• Mondays are not unprecedented, just rare — the last Monday entry in the entire changelog is DeepSeek V3.2 on 1 December 2025, roughly ten months before today.

• The holiday rule is written into the price sheet — DeepSeek's peak billing windows are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, "excluding Chinese public holidays," with all other hours off-peak including holidays in full. A company that carves public holidays out of its own billing calendar is telling you it does not operate them. National Day is 1 October.

That is the reading the tweet's second sentence lands on, and it is the more useful half. "At least next week until Thursday" pulls in the opposite direction from "on Monday": it pushes the plausible slot away from the day the tweet names and toward the boundary at the end of the working week, which is exactly where the holiday starts. Whichever Monday the author means, the Thursday that bounds the window is the last ordinary working day before the break. DeepSeek has not shipped anything through a National Day period.

Where a real release would show up first

This is the part worth bookmarking, because every one of these is checkable in one page load and none of them depends on a tracker's judgement:

• A second entry in the changelog. DeepSeek's API update page is the first place a model release appears, and it is where the two currently callable models are defined. Its newest entry today is the 10 September Flash release.

• A repository in the deepseek-ai organisation with Pro in the name. The organisation's newest repository is DeepSeek-V4.1-Flash from 10 September; a new one would be visible immediately.

• A third row in the harness's model catalogue. The adapter file above is two entries long, in a public repository, and a Pro tier the harness was tuned for would be the obvious addition.

• A price row and a concurrency limit. DeepSeek states availability limits on the price sheet, and a flagship tier would arrive with both.

• A tag above 0.1.x on npm or the desktop feed. For the harness half of the claim, this is the whole test — and the feeds have been static at 0.1.7-rc.2 since 24 September.

What to do in the meantime

There is a practical asymmetry in this story. The model everyone is waiting for has nothing to call, while the two DeepSeek models that do work are both routing today — and one of them, DeepSeek V4.1 Flash, is the release that actually moved the family forward two and a half weeks ago.

Both are on OrcaRouter behind one API key, at the provider's list price passed through with no markup: DeepSeek V4.1 Flash at $0.15 input and $0.60 output per million tokens and DeepSeek V4 Pro at $0.66 and $1.98, off-peak, with a 2× multiplier inside DeepSeek's peak windows. Because the pass-through is literal, DeepSeek's own peak and off-peak multiplier is what lands on the invoice, and a vendor-side price change appears on our side the day DeepSeek makes it rather than whenever a rate sheet is updated. If you are pinning a production path to one of these models, an automatic failover chain is the cheap protection against a provider-side wobble during a launch week — the kind of week when DeepSeek's API is most likely to be busy and a lot of people are rewriting their configs.

To be explicit about the limit: DeepSeek V4.1 Pro is not something we host, and we will not claim otherwise. It has no identifier to route. Nothing in this piece is a preview of it, and the moment it has a price row and a repository, it is a model-page problem rather than a rumour problem.

The honest summary

Two of the three assertions in this tweet are false as of today, in the strict sense that the artefacts they require do not exist. There is no 0.2 in DeepSeek Harness — 27 npm versions, 22 GitHub prereleases, two desktop feeds, all of them on 0.1.7-rc.2, all of them since 24 September. There is no DeepSeek V4.1 Pro — not in the changelog, the price sheet, the Hugging Face organisation, the harness's own model catalogue, or on our pages or Artificial Analysis's. The third assertion, that the next big thing is being trained for agent teams, is the one with real published work underneath it: a systems paper on DeepSeek's agent sandbox platform from 19 September, an Agent Teams plugin line that has been in the open since 9 September and is still switched off by default in the plugin page, and a model card that says out loud that DeepSeek's own agentic benchmark numbers were produced inside DeepSeek Harness. None of that is a schedule. All of it is why the schedule keeps being guessed at.

Treat the date as a hypothesis and the artefacts as facts. The changelog, the Hugging Face organisation and the harness's model catalogue are where this resolves, and all three are one page load away. Until one of them moves, a "together with V4.1 Pro on Monday" is a wish in a sentence that hedges itself to Thursday — and the Thursday in question is the last working day before a national holiday.