A Chinese lab just shipped a model that reads images, writes code, and drives agents — then priced its input tokens at the floor of the market. The strange part is the architecture: most of its 552 billion parameters never wake up. DeepSeek's V4.1-Flash, released 10 September 2026, is less a bigger model than a smaller one wearing a much larger coat.
What actually shipped
DeepSeek published V4.1-Flash on its official channels on 10 September 2026. It is the first model in a new architecture family, and it is open weight: the checkpoint sits on Hugging Face under an MIT licence, with a technical report beside it. The National Supercomputing Internet followed on 11 September with its own API and weight hosting, giving domestic developers a second, in-country path to run it.
On the API the model answers to the name deepseek-flash and becomes the default. The older V4-Flash and the separate vision experiment, V4-Flash-Vision-Exp, are retired — their names still resolve, but they now route to V4.1-Flash. DeepSeek had floated retiring the premium V4-Pro tier; after user pushback it reversed course and kept V4-Pro available past 14 September at unchanged billing. So the cheap model is now the front door, and the expensive one stays for whoever asks.
Native vision, finally
Previous Flash models needed a bolted-on vision encoder to read an image. V4.1-Flash trains text and images together from the start, so vision is native. For a product line whose entire reason to exist is low latency, that is a direction, not a bullet: DeepSeek is putting sight in the budget tier, not the premium one. The immediate use is practical — screenshot a form, paste a chart, let the agent act on what it sees — rather than a separate "multimodal" menu.
Why "asymmetric" is the whole story
The headline number — 552 billion total parameters — is the least useful thing to know. The split is what matters. DeepSeek describes a "Causal Encoder–Decoder" design: about 8 billion parameters activate to read your prompt, and about 16 billion to write the answer, out of the full 552B mixture of experts.
Reading is cheap, generating is dear, and almost every real workload is 90% reading. An agent that ingests a 200-page manual, a codebase, or a day of logs spends its tokens on input. By keeping the encoder half small, DeepSeek pushes the costly decoder work onto only the tokens it actually produces. That is a direct hit on the part of an AI bill that used to be invisible: re-reading the same context thousands of times.
The 890-byte trick
Long-context inference lives or dies on the KV cache — the memory that remembers what was already read. DeepSeek reports V4.1-Flash needs about 890 bytes of KV cache per token, roughly a quarter of the previous Flash generation and an eighth of its SSD footprint. Multiply across a million-token context held open for hours, and the difference is "we could never afford this" versus "we can leave it running."
The context window itself is 1 million tokens, with completions up to 384,000 tokens. For anyone self-hosting the open weights, the 890-byte figure — not the 552B headline — is what decides how many long sessions fit on a GPU card.
Open weights, with a reality check
MIT means anyone may download and fine-tune the model commercially. The honest footnote is size: the checkpoint runs to hundreds of gigabytes, so "open" does not mean "runs on your laptop." It means open to anyone with a server — which is most of the enterprises that would actually deploy it. Day-one support in open inference frameworks was widely reported, and that matters more than the licence text: an open model nobody can run is a PDF, not a tool.
What the benchmarks claim
DeepSeek's own launch table puts V4.1-Flash above its prior V4-Pro on agentic and coding tests — GPQA Diamond 90.9, Terminal-Bench 2.1 at 90.6, Automation-Bench 54.8, a Codeforces rating of 3471. On harder, newer reasoning tests the picture is mixed, and these are company-published numbers with no independent evaluation yet. Read them as DeepSeek's claims, not a verdict.
Pricing: the floor
The API keeps DeepSeek's peak/off-peak schedule. Off-peak, input is US$0.15 per million tokens (cache miss) and US$0.003 per million on a cache hit, output at US$0.60. Peak hours roughly double those. A cached-input rate of three hundredths of a cent is among the lowest on the market, and it rewards exactly the repetitive, context-heavy work the new architecture is built for.
What readers can do now
- If you build agents: prototype on
deepseek-flashand measure your bill against cache-hit rate — the win is largest when your context repeats. - If you self-host: budget GPUs against the 890-byte-per-token cache, not parameter count; that decides concurrent long sessions.
- If you bet on open models: wait for independent agentic benchmarks before trusting the launch table — the architecture is new and the scores are DeepSeek's own.
Honest limitations
Release date, MIT licence, Hugging Face weights, native vision, the retired model names, V4-Pro's continued availability, and the USD price schedule come from DeepSeek's official news page, API changelog, and pricing page (10 September 2026). The 552B total, 8B/16B active split, 890-byte KV cache, 1M-token context, training-token count, and all benchmark scores are DeepSeek's own published claims in its launch materials and model card; we found no independent third-party evaluation at writing and present them as company-reported. Checkpoint-size and self-hosting constraints are reported by third-party analyses, not DeepSeek. Parameter and cache figures can shift if DeepSeek revises its card.