DeepSeek V4.1 Flash shipped on September 10, 2026 as an MIT-licensed open-weights model: 552 billion total parameters, roughly 8 billion active at prefill and 16 billion at decode, and a 1 million token context window. Input runs $0.30 per million tokens at peak and $0.15 off-peak. The KV cache fell about 75%, to roughly 890 bytes per token. DeepSeek retires V4 Pro into Flash pricing on September 14.
DeepSeek did not launch a bigger model. It launched a cheaper one that beats its own flagship on most of the work people actually pay for.
That is the part that moves money. A frontier-adjacent model with permissive weights and a memory footprint one quarter the size of its predecessor changes what an inference dollar buys — for DeepSeek, and for everyone selling tokens against it.
What is DeepSeek V4.1 Flash?
DeepSeek V4.1 Flash is a 552-billion-parameter Mixture-of-Experts model released September 10, 2026 under the MIT license, distributed on Hugging Face and served through DeepSeek’s API as deepseek-flash. It activates about 8 billion parameters during prefill and 16 billion during decode, handles 1 million tokens of context, and accepts both text and image input.
The sparsity ratio is the headline. A 552B backbone that fires 8B parameters to read your prompt is doing roughly 1.4% of the arithmetic a dense model of that size would need.
Crypto Briefing reports the architecture uses an asymmetric causal encoder-decoder design, with the decoder reading from a global KV cache projected out of the encoder’s final hidden states.
BenchLM lists the model as text-and-image in, text only out, with an explicit reasoning mode that can be toggled per request.
The specification sheet
| Attribute | DeepSeek V4.1 Flash |
|---|---|
| Release date | September 10, 2026 |
| Total parameters | 552B (Mixture-of-Experts) |
| Active parameters | ~8B prefill / ~16B decode |
| Context window | 1,000,000 tokens |
| License | MIT |
| Modalities | Text + image in, text out |
| KV cache | ~890 bytes per token |
| API model name | deepseek-flash |
How much does DeepSeek V4.1 Flash cost?
Peak pricing is $0.30 per million input tokens and $1.20 per million output tokens, according to BenchLM’s tracked listing. Off-peak, those fall to $0.15 and $0.60. Cache hits are the outlier: $0.006 per million at peak and $0.003 off-peak, per Flowtivity’s breakdown of the pricing table.
A cache hit at $0.003 per million tokens is effectively free. That is the number to build architecture around.
DataStudios modeled a representative workload of 100 million input tokens and 10 million output tokens. Off-peak, the bill drops from roughly $28.60 to $21.00 — a 26.6% cut on like-for-like usage.
Where the discounts land
DataStudios breaks the off-peak reductions down by token type:
- Cached input: $0.003 per million, a 57.1% reduction
- Uncached input: $0.15 per million, a 31.8% reduction
- Output: $0.60 per million, a 9.1% reduction
The asymmetry is deliberate. DeepSeek cut hardest where its own new architecture saves it the most — cached context — and barely moved output, where it still burns real compute.
For teams running long-context agent loops that re-read the same repository or document set on every turn, that pricing shape is worth more than the headline rate.
How much better is it than V4 Pro?
On agentic and coding work, substantially. DataStudios reports V4.1 Flash scoring 30.0 on Terminal-Bench 3.0 against V4 Pro’s 11.8, and 74.2 on DeepSWE against V4 Pro’s 62.7. On pure reasoning, V4 Pro still edges ahead: GPQA Diamond of 92.4 versus Flash’s 90.9.
DeepSeek’s own API documentation states that “V4.1 Flash has comprehensively surpassed V4 Pro in performance, cost, speed, and total time.”
That sentence is a retirement notice dressed as a benchmark claim.
V4 Pro is being switched off in everything but name
From September 14, 2026, DeepSeek reroutes V4 Pro API requests to V4.1 Flash and bills them at Flash rates until a V4.1 Pro arrives. The company confirms this directly in its API documentation.
Legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp keep working, routed to current models at updated pricing.
DataStudios calculates the swap can cut some workloads by up to 75.5%, because customers are not just getting a discount — they are being moved onto a different, cheaper model.
Why does the KV cache number matter to investors?
Because KV cache is what actually fills the memory on an inference server. Crypto Briefing reports V4.1 Flash holds roughly 890 bytes per token against about 3,560 bytes for V4-Flash — a 75% reduction that lets the same hardware serve roughly four times as many concurrent users, or contexts four times as long.
Serving capacity is the constraint that has been repricing the entire AI stack. Memory, not FLOPs, is what runs out first at long context.
That is the same shortage now pushing Chinese accelerator prices up: Huawei’s Ascend 950DT crossed 250,000 yuan, roughly $37,000, up 20% to 50%, with Cambricon’s 690 up 20% to 30%, driven by high-bandwidth memory scarcity, per reporting compiled by Tech Startups on September 10.
A model that needs a quarter of the memory per token is a hedge against exactly that squeeze. It is also why open weights matter here — anyone renting GPUs can capture the saving directly rather than paying an API markup for it.
Who wins and who loses financially?
Inference resellers and GPU-rental businesses win, because a permissively licensed model with a small memory footprint raises their gross margin without any new hardware. Closed-model vendors selling mid-tier tokens lose pricing power. DeepSeek itself trades revenue per token for volume.
The MIT license is the aggressive part. There is no royalty, no usage gate, no acceptable-use clause to negotiate with a procurement team.
That puts direct pressure on the cheap-tier products from every lab that charges for equivalent capability. We covered the same dynamic when Tencent dropped Hy4 as 770B open weights, and again in the GLM-5.3-Flash and Qwen3.8-Flash-Next matchup.
The long-context angle
A million-token window at $0.15 per million input tokens off-peak resets the floor on long-context pricing. That market was already compressing — see our look at the cheapest 1M context model — and a free-to-self-host entrant compresses it further.
The catch is that self-hosting 552B parameters is not cheap either. The weights are free; the servers are not.
Is DeepSeek V4.1 Flash actually frontier-class?
Not across the board, and the gaps are specific. Flowtivity reports Claude Opus-5.0 leading Terminal-Bench 4.0 by 51.8 to Flash’s 31.2 — a wide margin on long-horizon, repository-scale work. BenchLM lists Flash at 20.3 on ProgramBench.
Short agent loops are where it competes. Long ones are where it does not.
The comparison DeepSeek’s supporters lead with is DeepSWE v1.1, where Flowtivity reports Flash at 74.2 against GPT-5.6 Sol at 73.0 and Claude Opus-5.0 at 74.0. That is a statistical tie at the top, not a win.
The skeptical read
BenchLM lists V4.1 Flash as tracked but not publicly ranked, citing insufficient non-generated benchmark coverage for leaderboard inclusion. Of its 22 source-displayable benchmark rows, reasoning, multilingual and instruction-following show no measured data at all.
Most of what we know about this model 24 hours in came from the party selling it.
There is also the provenance question hanging over every Chinese frontier release right now, after the distillation advisory naming six Chinese firms over token theft. Nothing in that advisory names DeepSeek’s V4.1 line, and no evidence has been presented tying this release to it — but enterprise buyers are asking the question, and that friction has a cost.
Benchmark table
| Benchmark | V4.1 Flash | Comparison | Source |
|---|---|---|---|
| DeepSWE v1.1 | 74.2 | Claude Opus-5.0 74.0 / GPT-5.6 Sol 73.0 | Flowtivity |
| Terminal-Bench 3.0 | 30.0 | V4 Pro 11.8 | DataStudios |
| Terminal-Bench 4.0 | 31.2 | Claude Opus-5.0 51.8 | Flowtivity |
| GPQA Diamond | 90.9 | V4 Pro 92.4 | DataStudios / BenchLM |
| CyberGym | 88.1 | — | BenchLM |
| BabyVision | 89.6 | — | BenchLM |
| ProgramBench | 20.3 | — | BenchLM |
What should a buyer do this week?
Test it on the narrow, repetitive agent work first, because that is where the evidence is strongest and the cache pricing pays off fastest. Do not migrate long-horizon repository tasks until independent Terminal-Bench 4.0 numbers land.
A practical sequence:
- Benchmark your own agent loop against V4.1 Flash off-peak, where input is $0.15 per million.
- Restructure prompts so the stable context hits cache at $0.003 per million rather than being re-sent uncached.
- Audit anything still calling
deepseek-v4-probefore September 14, when it silently becomes Flash. - Price a self-hosted deployment against API rates — MIT weights only pay off above a certain volume.
- Hold long-horizon coding on whatever currently wins your internal evals.
Frequently asked questions
When was DeepSeek V4.1 Flash released?
September 10, 2026, on Hugging Face and through the DeepSeek API under the model name deepseek-flash.
Is DeepSeek V4.1 Flash open source?
The weights are published under the MIT license, which permits commercial deployment without royalties. Training data and training code are a separate question and are not covered by that license grant.
How many parameters does DeepSeek V4.1 Flash have?
552 billion total, with roughly 8 billion active during prefill and 16 billion during decode. It is a Mixture-of-Experts model, so total and active counts differ sharply.
What happens to DeepSeek V4 Pro?
From September 14, 2026, V4 Pro API requests are rerouted to V4.1 Flash and billed at Flash pricing, until a V4.1 Pro is released. DeepSeek states this in its API documentation.
How much does DeepSeek V4.1 Flash cost per million tokens?
$0.30 input and $1.20 output at peak; $0.15 and $0.60 off-peak. Cache hits are $0.006 at peak and $0.003 off-peak.
Does DeepSeek V4.1 Flash beat GPT-5.6 Sol?
On DeepSWE v1.1, Flowtivity reports 74.2 against 73.0 — close enough to call even. On Terminal-Bench 4.0 it trails Claude Opus-5.0 by 31.2 to 51.8, so the answer depends entirely on the task length.
What is the context window?
One million tokens, with image input supported alongside text.
The bottom line
DeepSeek V4.1 Flash is not a frontier model and does not need to be. It is a margin weapon: 552B of capability with an 8B activation cost, MIT-licensed, at $0.15 per million input tokens off-peak and a KV cache small enough to quadruple what a server can hold.
The verdict is narrow but real. For short agentic loops and cache-heavy long-context work, this is the cheapest serious option shipped this month. For repository-scale reasoning, it is not close to the leaders, and DeepSeek’s own GPQA number went backwards.
The number that matters most is not on any leaderboard. It is 890 bytes per token — because that is what determines how many customers a given rack of GPUs can serve, and therefore who can afford to sell tokens at these prices without losing money on every call.
Treat the benchmark spread as provisional until independent evaluators publish. One day of vendor-supplied numbers is not a track record.
Sources
- DeepSeek API Documentation — model routing and V4 Pro transition
- BenchLM — DeepSeek V4.1 Flash benchmarks and pricing
- Crypto Briefing — DeepSeek launches V4.1-Flash with 552B parameters
- Flowtivity — DeepSeek V4.1 Flash benchmark analysis
- DataStudios — pricing changes and V4 Pro replacement
- Tech Startups — September 10, 2026 tech news roundup
Leave a Reply