DeepSeek V4.1 Flash cuts long-context memory

In an agent, reading usually takes far more space than writing. The model receives instructions, files, tool results, and perhaps hundreds of thousands of tokens from earlier attempts; it then produces a fraction of that volume. A conventional transformer nevertheless preserves, for every layer, attention state associated with the full context. The longer the session and the greater the concurrency, the more memory is devoted to remembering before generating the next token.

DeepSeek released DeepSeek‑V4.1‑Flash on September 10, an open-weight multimodal model that turns this usage imbalance into an architectural one. This is not another post-training pass over V4‑Flash, like the 0731 checkpoint published in July. It is a model trained from scratch with a 552-billion-parameter backbone, native vision, and a stated one-million-token window. The most important novelty is not that it stores more parameters, but that it tries to move much less state and compute through the most repeated portion of an agent workload.

Read once, reuse many times

The V4.1‑Flash technical card describes 40 layers divided into two halves: a causal encoder processes the input and a decoder generates the output. Rather than building a separate global key-value cache in every decoder layer from different hidden states, the system projects it from the encoder's final representation. It therefore activates about 8 billion parameters per token during the initial read and 16 billion during generation. The previous V4‑Flash activated 13 billion in both phases.

That difference does not make V4.1 an 8-billion-parameter model. The full backbone still has 552 billion parameters and adds 196 billion of Engram conditional memory, accessed sparsely. “Active parameters” approximates the work a token traverses, not the storage required to own the model or all the communication required to serve it.

DeepSeek combines that separation with three memory techniques. Sparse attention shares state and indices across layers; a hierarchical index reduces the candidates later layers must examine; and part of the cache is quantized to four bits. For sliding-window layers, the model retains recent tokens and reconstructs state when needed instead of persisting all of it to SSD. According to the lab's report, the global cache is 890 bytes per token: roughly one quarter of the high-bandwidth memory and one eighth of the persistent storage required by V4‑Flash.

This is a systems improvement, not proof of perfect understanding across a million tokens. A smaller cache can fit more sessions or longer histories on the same hardware; it does not establish that the model retrieves every distant fact, preserves a plan throughout the window, or avoids having irrelevant material degrade a decision. Nor is there yet an outside measurement isolating the contribution of the encoder-decoder design, sparse attention, and quantization. The memory figures come from DeepSeek, although the weights and reference code make them testable.

The architecture also incorporates images from pre-training: a vision encoder converts each image into representations that mix with text from the beginning. The declared corpus contains 45 trillion multimodal tokens. This replaces August's separate experimental vision model with one endpoint that can combine screenshots, documents, and tools—a practical difference for agents operating interfaces or repositories containing diagrams.

Capability, cost, and openness that demands machines

DeepSeek presents its agent gains as evidence that efficiency did not diminish capability. In its own runs, V4.1‑Flash solves 90.6% of Terminal‑Bench 2.1 versus 82.7% for July's V4‑Flash, and 74.2% of DeepSWE v1.1 versus 54.4%. The first evaluation measures verifiable tasks in a terminal; the second measures software changes against tests. But both figures use maximum effort, as many as 500 steps, and multiple samples per task. They measure a model inside a harness, not a single call or the expected cost of repairing an arbitrary repository.

The evidence has improved since July. The repository now includes pinned instructions for reproducing DeepSWE with both DeepSeek's minimal agent and mini‑swe‑agent, and DeepSeek Harness is published under MIT. A full external reproduction of the figures is still missing, but the exact scaffold behind one central claim no longer has to be guessed.

A second signal, independent of the lab, helps keep that result from becoming a general claim. In the September 16 snapshot of Agent Arena, which derives signals from real sessions, V4.1‑Flash ranks thirteenth among 45 models: well behind the leading frontier models on the combined score, but third in how often a user confirms that the task is complete. Its observed median cost is six cents per task, compared with several dollars near the top. That is strong evidence of an interesting cost-utility frontier; it is not a controlled experiment or a promise of success under another mix of users, tools, or prices.

Openness makes this difference investigable. DeepSeek publishes the weights under MIT, the message format, readable inference code, the image processor, and the architectural components. The vLLM community already maintains a model-specific serving recipe. But “open weights” once again does not mean “local model”: the minimal implementation uses eight-way process parallelism in its example, and the announcement itself invites organizations with roughly 2,000 GPUs and a storage cluster to discuss large-scale deployment. A small organization is more likely to inspect open artifacts while consuming the model through an API.

That API makes the design's economics concrete. Current pricing is $0.30 per million uncached input tokens and $1.20 per million output tokens at peak time; off-peak costs half as much, while cached input falls to $0.006 or $0.003. It is easy to see why DeepSeek optimizes reading and retention first. But token price is not outcome price: maximum effort may generate tens of thousands of tokens, repeat tools, or finish without solving the task.

Builders also need to pin the identifier, not just the commercial name. Calls to the old deepseek-v4-flash and deepseek-v4-flash-vision-exp identifiers are now routed to V4.1‑Flash. The original announcement said deepseek-v4-pro would follow on September 14; the updated change log now says DeepSeek will keep V4‑Pro available in response to users. A compatible endpoint can change its underlying weights while client code stays untouched.

My reading is that V4.1‑Flash matters less for topping a table than for proposing a different unit of efficiency. Until now, a “cheap model” usually implied fewer parameters or less reasoning effort. Here it means spending differently when reading, remembering, and writing. If external measurements confirm the memory reduction, that idea could expand both the number and duration of concurrent agents without making them trim their history at every step.

Two important tests remain. The first is operational: measure throughput, energy, degradation on real contexts, and cost per solved task on outside installations. The second is safety. The card and report publish cyber results, but no V4.1-specific evaluation of misuse, refusal, or action control. In an MIT-licensed model designed for agents and tools, a fourfold smaller cache widens deployment; it does not replace permission limits, isolation, or a public characterization of failure.

Sources

• DeepSeek, DeepSeek‑V4.1‑Flash: Smarter, Faster, More Efficient, September 10, 2026.
• DeepSeek‑AI, DeepSeek‑V4.1‑Flash card, weights, code, and technical report, accessed September 16, 2026.
• DeepSeek, API models and pricing and change log, accessed September 16, 2026.
• DeepSeek‑AI, DeepSeek Harness, development version accessed September 16, 2026.
• Arena Intelligence, Agent Arena, September 16, 2026 snapshot.
• vLLM, DeepSeek‑V4.1‑Flash deployment recipe, accessed September 16, 2026.