DeepSeek released DeepSeek-V4.1-Flash on 10 September 2026, and the release does three things at once: it ships a 552B-parameter multimodal Mixture-of-Experts model as MIT-licensed open weights on Hugging Face, it cuts cache-hit API pricing by 57%, and it retires two shipping models on the spot. DeepSeek-V4-Flash and DeepSeek-V4-Flash-Vision-Exp are gone today. DeepSeek-V4-Pro, the 1.6T flagship the company shipped in August, starts routing to V4.1-Flash at 04:00 UTC on 14 September. If you have deepseek-v4-pro hardcoded anywhere, you have four days before a different model answers that call.
The headline the company chose is about KV cache compression, which sounds like an infrastructure footnote. It is not. The compression is the reason the prices moved, and the prices are the part that lands on your invoice.
What DeepSeek Shipped
V4.1-Flash is a Causal Encoder-Decoder model: 40 transformer layers split into a 20-layer causal encoder and a 20-layer decoder, where the decoder's global KV cache is projected from the encoder's final hidden states rather than derived layer by layer. That structure is why the model activates only 8B parameters per token during prefill and 16B during decode, despite a 552B backbone. There are 384 routed experts plus one shared expert, with six routed experts active per token.
The compression numbers are specific. The global KV cache is 890 bytes per token, roughly a quarter of DeepSeek-V4-Flash and, by the company's own accounting in the technical report, about 437 times smaller than DeepSeek-V1. Three mechanisms get it there: Compressed Sparse Attention 2, which assigns each attention layer a Full, Reindex, or Reuse mode so that only four of the forty layers actually source main KV; FP4 caching of that main KV in E2M1 format with one E4M3 scale per 16 channels; and SWA Bounded Replay, which reconstructs sliding-window KV state by replaying recent tokens instead of persisting it, cutting SSD footprint to an eighth.
Vision is native this time, not a side experiment. A from-scratch DeepSeek-ViT encoder with 2D-RoPE and 3x3 pixel-unshuffle downsampling feeds visual embeddings into the language model from the start of pre-training, on a 45T-token multimodal corpus. Context is 1M tokens, extended from a 64K training length via YaRN, with a maximum output of 384K.
The Price Cut, Measured Against the Model It Replaces
DeepSeek's announcement says prices are lower without saying lower than what. Here is the comparison, taken from the current pricing page and from an archived snapshot of the same page from 1 September, before any of this shipped. All figures are US dollars per 1M tokens, shown as off-peak / peak.
| Rate | V4-Flash-0731 (retired today) | V4-Pro-0813 (rerouted 14 Sep) | V4.1-Flash (now) |
|---|---|---|---|
| Input, cache hit | $0.007 / $0.014 | $0.022 / $0.044 | $0.003 / $0.006 |
| Input, cache miss | $0.22 / $0.44 | $0.66 / $1.32 | $0.15 / $0.30 |
| Output | $0.66 / $1.32 | $1.98 / $3.96 | $0.60 / $1.20 |
| Context | 1M | 1M | 1M |
| Native vision | No, separate Exp model | No | Yes |
| Concurrency limit | 2,500 | 500 | 2,500 |
Against the Flash model it replaces, the cut is uneven and deliberately so: cache-hit input drops 57%, cache-miss input drops 32%, and output drops only 9%. DeepSeek is explicit about why. Cache-hit charges dominate agent bills, because an agent re-sends a growing conversation on every turn and almost all of it is a repeat. Compressing the cache is what made that specific line item cheap, so that is the line item that fell hardest.
Against V4-Pro, the model being phased out, the gap is much larger: 86% off cache-hit input, 77% off cache-miss input, and 70% off output. Concurrency goes from 500 to 2,500 at the same time. That is the real story for anyone who chose Pro for quality.
Your V4-Pro Calls Get Rerouted on 14 September
This is the part that can break a working pipeline, and it is buried in a footnote. From 04:00 UTC on 14 September 2026, every request to deepseek-v4-pro is served by V4.1-Flash and billed at V4.1-Flash rates, and DeepSeek says this continues until a V4.1-Pro exists. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp already route to V4.1-Flash as of today.
Your bill goes down. Your model changes. Those are not the same event, and only one of them is announced as good news. If you have prompts tuned against V4-Pro's behaviour, evaluation suites pinned to its outputs, or a golden-transcript regression test, run them against deepseek-flash this week rather than discovering the difference on Monday. The new model also uses a different sampling default: DeepSeek recommends temperature=1.0 and top_p=0.95, with a controllable reasoning effort expressed as an integer from 1 to 100 that trades cost against accuracy. Every benchmark DeepSeek published used effort 100, which is the expensive end.
One more compatibility detail worth checking before Monday: the concurrency limit attached to your account changes with the model, from 500 on Pro to 2,500 on Flash, and the rate limit documentation is where the isolation rules for that live.
The Peak-Hours Detail That Halves Your Bill
DeepSeek prices peak and off-peak, with off-peak set at exactly half of peak. Peak is 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday. Everything else, including all weekend, is off-peak.
Convert that to Eastern Time and something useful falls out. Peak runs 9 PM to midnight ET, and 2 AM to 6 AM ET. Both windows sit entirely outside a normal North American working day. A creator in New York or Toronto who runs batch jobs between 9 AM and 6 PM is already paying the off-peak rate on every call, without configuring anything. The scheduling advice DeepSeek gives, which is to move flexible workloads off-peak, is advice for users in Asian time zones. For a US-based creator the pricing table's off-peak column is simply the price, and the numbers to plan against are $0.003 per 1M cached input tokens and $0.60 per 1M output tokens.
Where the Benchmarks Are Strong, and Where They Are Not
DeepSeek published a comparison against Opus-5.0, GPT-5.6 Sol, K3, GLM-5.3, and its own V4-Pro and V4-Flash. Read top to bottom, it splits cleanly into two halves.
V4.1-Flash leads the table on Terminal-Bench 2.1 (90.6 against Opus-5.0's 89.1), DeepSWE v1.1 (74.2 against 74.0), AutomationBench (54.8 against 50.3), Agent's Last Exam (31.8 against 28.6), CyberGym (88.1), and Codeforces, where its 3471 rating beats V4-Pro's 3348. For an 8B-active model against frontier flagships, that is a genuinely striking result.
Then look at the harder benchmarks. On Terminal-Bench 3.0 it scores 30.0 against Opus-5.0's 43.3. On Terminal-Bench 4.0 it scores 31.2 against 51.8. On Humanity's Last Exam without tools it scores 36.8 against 56.3. On ProgramBench, 20.3 against 37.0. On NL2Repo-Bench, 64.0 against 75.3.
The pattern is not random. Terminal-Bench 2.1 is close to saturated, with every model in the table between 82.7 and 90.6, so a 1.5-point win there separates almost nothing. Terminal-Bench 4.0 still spreads the field across 40 points, and on that spread the frontier models are ahead by a wide margin. The honest read, which matches what early technical coverage emphasises, is that V4.1-Flash has closed the gap on the class of agentic work that is already routine and has not closed it on the work that is still hard. At roughly a fifth of Pro's price, that trade is often the right one. It is still a trade.
The Harness Changes the Score More Than You Think
DeepSeek also published something most labs do not: the same model's scores across eight different agent scaffolds. On DeepSWE v1.1, V4.1-Flash resolves 74.2% under mini-SWE, 72.6% under DeepSeek Harness in Minimal mode, 69.8% under Claude Code, 66.2% under Pi, 65.6% under Codex, and 65.5% under OpenCode. On Terminal-Bench 2.1 the spread runs from 90.6% down to 84.1%.
That is an 8.7-point swing on DeepSWE from harness choice alone, on one fixed model. It is larger than most of the model-to-model gaps in the headline comparison table. It also means the 74.2 quoted in the flagship row was measured under mini-SWE, the scaffold that happened to score best, while a developer running the same model inside Claude Code should expect closer to 69.8.
None of that is dishonest, and publishing the full scaffold table is more transparency than the field norm. But it is a reminder to benchmark inside your own harness before you switch a production pipeline on the strength of a leaderboard row.
What Open Weights Means at 510 GB
The MIT license is real, and it is the most permissive license on any model at this capability tier. The practical barrier is size. The Hugging Face repository holds 48 safetensors shards totalling roughly 510 GB, about 475 GiB on disk, and that is the already-quantised artifact: FP8 weights with FP4 experts, 32x32 blocks, ue8m0 scale format. On top of the 552B backbone sits a further 196B parameters of Engram conditional memory, a sparse token-lookup memory with a 16M-entry vocabulary attached at layers 1 and 14.
DeepSeek does not pretend otherwise. Its own announcement closes by inviting anyone "planning a large-scale deployment with 2,000 GPUs and a storage cluster" to get in touch. That is the honest scale of self-hosting this. Open weights here means auditable, forkable, and free of vendor lock-in at the licence level. It does not mean you can run it on a workstation, and any coverage suggesting otherwise has not looked at the file listing.
Two rough edges are worth knowing before you try. The release ships no Jinja chat template, so prompt construction goes through the reference Python implementation in the repository's inference folder or through the encoding module beside it. The model card also points to a Rust toolkit called deepseek-recipe for production prompt encoding, and that repository returns a 404 at the time of writing, so treat it as announced rather than available.
How to Try It Today
The fastest path costs nothing structural. The API model name is deepseek-flash, and DeepSeek serves both an OpenAI-compatible endpoint at https://api.deepseek.com and an Anthropic-format endpoint at https://api.deepseek.com/anthropic. Any tool that already speaks either protocol needs a base URL and a model string, not a rewrite.
- Point an existing OpenAI or Anthropic SDK client at the matching DeepSeek base URL and set the model to
deepseek-flash. - Start reasoning effort low. The scale runs 1 to 100 and the reasoning effort guide documents the trade. Every published benchmark used 100, so treat that as the ceiling rather than the default.
- Send an image. Native vision is now in the cheap tier rather than a separate experimental model, and images are billed as input tokens by dimension, which makes screenshot and document pipelines dramatically cheaper than they were last week.
- If you would rather not manage a second API key, the model is already listed on OpenRouter at $0.30 in and $1.20 out per 1M tokens, which is the peak rate. Going direct is cheaper for most of the day.
- Re-run your own evaluations before 14 September, because that is when V4-Pro traffic moves.
What This Does to the Open-Weights Price Floor
Six weeks ago DeepSeek's open-weight coding model and its 1.6T flagship were separate products at separate prices, and adding vision meant calling a third, experimental endpoint that we covered as vision at no price premium. Today all three are one model, it is multimodal by default, it costs less than the cheapest of them did, and the flagship is being retired into it.
The competitive pressure lands on everything in the same tier. Our three-way comparison of GLM-5.3-Flash, Qwen3.8-Flash-Next and DeepSeek V4 was priced against a $0.22 cache-miss floor. That floor is now $0.15, with 86% off cached input, and the model behind it carries an MIT licence and native vision. Every open-weights lab shipping into that band has to answer it, and the answer is unlikely to be a price rise.
The caution worth keeping is the one the benchmark table already told us. This is not a frontier-beating model on genuinely hard agentic work, and the Terminal-Bench 4.0 gap is 20 points wide. What it is, quite precisely, is the cheapest place to run the agentic work that has stopped being hard, with an image encoder attached and a licence that lets you leave.
Frequently asked questions
What happens to my deepseek-v4-pro API calls?
Nothing until 04:00 UTC on 14 September 2026. From that moment every deepseek-v4-pro request is served by V4.1-Flash and billed at V4.1-Flash rates, which is 77% less for cache-miss input and 70% less for output. DeepSeek says this continues until a V4.1-Pro launches. The model name keeps working, so nothing breaks at the transport layer, but a different model answers.
Is DeepSeek-V4.1-Flash actually cheaper than DeepSeek-V4-Flash?
Yes, across every rate, though not evenly. Cache-hit input falls 57% from $0.007 to $0.003 per 1M tokens off-peak, cache-miss input falls 32% from $0.22 to $0.15, and output falls 9% from $0.66 to $0.60. The cache-hit cut is the largest because compressing the KV cache is what the release is built around, and cached input is where agent workloads spend most of their money.
Can I run DeepSeek-V4.1-Flash on my own hardware?
Only on a cluster. The weights are MIT-licensed and freely downloadable, but the repository is 48 shards totalling about 510 GB, and that is already FP8 with FP4 experts rather than a full-precision checkpoint. DeepSeek's own release invites deployments of 2,000 GPUs plus a storage cluster. It is open in the licensing sense, not the run-it-on-a-desktop sense.
Does it handle images natively?
Yes. A purpose-built DeepSeek-ViT encoder was trained from scratch and integrated from the start of language-model pre-training, on a 45T-token multimodal corpus, with a cap of 1,024 image tokens per image. This replaces the separate experimental vision endpoint, which is retired as of today, and it costs less than that endpoint did.
How reliable are the benchmark scores?
Treat the strong rows as real and the framing as favourable. The wins cluster on near-saturated benchmarks such as Terminal-Bench 2.1, where the whole field sits within eight points, while the losses cluster on harder ones such as Terminal-Bench 4.0, where Opus-5.0 leads by more than 20 points. DeepSeek also published per-scaffold results showing an 8.7-point swing on DeepSWE from harness choice alone, and the headline number used the best-scoring harness.
What is the reasoning effort setting?
An integer from 1 to 100 that continuously trades inference cost against accuracy, adjustable per request. All of DeepSeek's published evaluation numbers were produced at 100, the most expensive setting, so real-world cost at a usable quality level depends heavily on where you set it. Start low and raise it only where a task demonstrably needs it.
