DeepSeek V4.1-Flash shipped under an MIT licence, a multimodal model of more than 550 billion parameters whose offloaded cache drops to an eighth of the previous generation. API pricing does not move, context reaches one million tokens, and the target is stated plainly: agents that run for hours.
Key Takeaways
- V4.1-Flash comes down to 890 bytes of cache per token, a quarter of V4-Flash in GPU memory.
- The model landed on September 10 on Hugging Face, open weights under an MIT licence.
- It edges past Opus 5 and GPT-5.6 Sol on DeepSWE v1.1 while openly declaring gaps elsewhere.
Have an AI Sum Up This Article
ChatGPTEight times less cache for the same running agent
One number carries the whole DeepSeek V4.1-Flash announcement: 890 bytes per token. That is what the model’s global cache now weighs, a figure the model card published on Hugging Face documents line by line, alongside the open weights and the model’s 48 shards.
Set against the previous generation, the cache held in GPU memory falls to a quarter of V4-Flash. The permanently offloaded portion, the one kept on disk so a prefix can be reused later, drops to an eighth. Measured against the house’s very first model, the ratio reaches 437.
Part of the gain comes from a format decision. The main cache moves to FP4 where the previous generation stayed on FP8, which mechanically halves the footprint before any architectural optimisation is counted. The topic is a sore one at DeepSeek, which raised its API prices in mid-August with cache reads six times more expensive.
The architecture itself steps away from the usual layout. DeepSeek V4.1-Flash runs a 40-layer causal encoder-decoder, split into a 20-layer causal encoder followed by a 20-layer decoder. That split lets it activate only 8 billion parameters while reading, against 16 billion while writing.
The asymmetry is not cosmetic, it targets one usage profile. An agent chaining tool calls spends most of its life reading fresh output rather than composing it, so halving the price of reading halves the price of the agent.
The problem it addresses is familiar to anyone running long agents. Agent tokens have jumped fourteen times since February, and the bill climbs mostly because an agent drags a history behind it that never stops growing.
Benchmarks held, gaps declared
On DeepSWE v1.1, DeepSeek V4.1-Flash posts 74.2 percent, which puts it slightly ahead of Opus 5 and GPT-5.6 Sol. On Terminal-Bench 2.1 it reaches 90.6 percent, and 54.8 percent on AutomationBench.
DeepSeek does not hide the other side of the sheet. The lab concedes persistent gaps on ProgramBench, on complex scientific tasks and on image analysis, three grounds where it makes no claim to match the closed frontier.
The training method is worth flagging. The model was trained from scratch on 45 trillion tokens of text and images, with a deliberate refusal to introduce new algorithmic methods. The bet rides on more data, better controlled, rather than on a technical discovery.
That continues a trajectory readable since the summer. V4-Flash-Vision was already closing in on Opus 4.8 in August, and the house advances through tight iterations rather than announced ruptures.
Context reaches one million tokens, and the model handles images natively alongside text. On paper the package covers most of what a production agent touches daily.
What remains untested is behaviour in the hand, where spec sheets stop being useful. Our V4-Flash-Vision test found genuine solidity on agent work, and that same ground is what needs remeasuring on this release.
More articles on Horizon
- ChatGPT Pro Has Stopped Taking New Subscribers
- Anthropic Modeled 17.9% Knowledge Worker Unemployment
- Paul Christiano Joins the Board That Clears OpenAI Models
A flat price that shifts the comparison
The commercial detail speaks louder than the benchmarks. The API keeps exactly the V4-Flash grid, so the drop in memory footprint is not billed back at the counter. For a team, that means cheaper agents to run with no renegotiation and no contract migration.
Holding the line cuts against the previous sequence, the August price rise followed by V4-Pro being pulled from testing. The house is climbing on capability while stepping back down on cost to run.
The MIT licence is the second lever. Genuinely open weights, with no field-of-use clause, allow execution on owned infrastructure, which neither OpenAI nor Anthropic offers on their frontier models.
For the closed labs the pressure does not land on the leaderboard, it lands on the cost structure. A competitor that makes agentic inference markedly cheaper, at comparable quality on code, forces the price gap to be justified by something other than raw performance.
The Chinese open-weight field is thickening fast, and memory is becoming the differentiator where the score was six months ago. Whoever keeps a long agent alive on less hardware wins the volume deployments.
The thread to pull over the coming weeks is pricing position. V4 Flash was already nearing GPT-5.6 Luna at 60 percent lower cost, and this release tightens the real cost gap further on long-running work.
Follow the story on Horizon.
