DeepSeek V4.1 Flash: Sept 10 Launch, Fast Uncensored Forks

A release date that got scrambled

DeepSeek V4.1 Flash shipped on September 10, 2026. It did not ship on October 4. The mix-up shows how fast the open-weight AI ecosystem moves.

The timeline is well documented. The Hugging Face repository went up at 02:17 UTC on September 10. DeepSeek's six-post announcement thread followed on X at 06:10 UTC, which was 14:10 Beijing time. A same-day entry appeared in the API change log. 3 That matched DeepSeek's own advance notice, which had pointed to around September 10 Beijing time. 3 A brief two-day beta came first, carrying the internal tag "deepseek-v4.1-flash-expires-on-0910." 2

The October 4 date comes from a model-tracking listing. It shows a "luminal/DeepSeek-V4.1-Flash" entry as released and added to that gateway on October 4, and calls it the only new model of October so far. 5 That listing points back to the original deepseek-ai repository. 5 The likeliest reading is that October 4 marks when a third-party host started serving the model, not when DeepSeek released it. It is a good example of why aggregator timestamps need checking against first-party change logs.

What the model actually is

DeepSeek calls V4.1 Flash the smallest model in a new architecture family. It describes that family as built for a higher capability ceiling, faster inference, higher throughput and scaling to larger models. 4

Key details:

• Architecture: a 552B mixture-of-experts model with about 8B parameters active. 3
• Vision: native visual understanding. 34
• Context and modes: a 1M-token context window, with both thinking and non-thinking modes. 1
• License: MIT, which allows free download, modification, fine-tuning and commercial redistribution. 23
• Price: lower Flash pricing than before. 3

The API naming changed as well. The model is served under a versionless ID, deepseek-flash. 34 The older V4 Flash and V4 Flash Vision Exp models were retired, and their legacy IDs now route to V4.1 Flash. 14

DeepSeek's own launch table claims V4.1 Flash beats V4 Pro on every agentic coding, cyber and tool-use row it reports, at a lower price. 1 These are vendor numbers. One benchmark aggregator gives V4 Pro a public score of 66/100 but has not yet ranked V4.1 Flash independently. 1 Until outside evaluations arrive, the "small model beats flagship" claim should be treated as promising but unverified.

The V4 Pro question

Accounts differ on what happened to V4 Pro. DeepSeek's announcement said V4.1 Flash outperforms V4 Pro on performance, cost, speed and total time. It said that after 12:00 Beijing time on September 14, all deepseek-v4-pro requests would route to V4.1 Flash at Flash prices, until a future V4.1 Pro arrives. 4 Some coverage reported this as a done deal, with V4 Pro retired and its traffic redirected. 2

A more recent model guide says otherwise. It reports that DeepSeek announced the reroute but then kept V4 Pro in service with no published end date. It lists both deepseek-flash and deepseek-v4-pro as current API IDs. 1 Developers on V4 Pro should check live API behavior rather than rely on the original announcement. The guide recommends V4 Pro only for existing integrations. 1

Jailbroken within hours

The most notable part of the story came after launch. "Abliterated" builds of V4.1 Flash appeared on Hugging Face within hours. These are modified versions with the model's refusal behavior stripped out. The most-tracked fork logged 2,254 downloads. 2 Tech Insider called it the fastest documented turnaround of the year from open-weight release to jailbroken fork. It noted that nearly a month later, none of the structural questions the episode raised had been answered. 2

This follows a repeatable pattern, as that report notes. 2 What makes V4.1 Flash a sharper test is how capable it appears to be. The MIT license makes the outcome close to inevitable: once weights are public under such terms, a modified fork needs no one's permission. The download count is modest in absolute terms. The real issue is the speed, which suggests safety tuning on open weights lasts for hours, not weeks.

The takeaway

V4.1 Flash is a meaningful release. It is a compact, vision-capable, long-context model that DeepSeek positions above its own previous flagship, at a lower price and under one of the most permissive licenses available. It is also a reminder of two tradeoffs that come with open-weight releases:

1. Dates are messy. Secondary trackers can mislabel when a model actually launched.
2. Guardrails are temporary. Fully open weights mean safety tuning can be removed almost immediately.

Anyone evaluating the model should verify against DeepSeek's change log, wait for independent benchmarks, and assume that uncensored variants exist alongside the official one.