DeepSeek V4.1 Flash brings 1M-token context, 384K output and cheap API pricing

What DeepSeek released

On September 10, 2026, DeepSeek released V4.1 Flash, an open-weight multimodal model positioned as the production successor to its earlier Flash line. According to a deep dive published on dev.to, the model accepts text and image input, produces text output, and ships under an MIT checkpoint licence. Its headline specifications are a one-million-token context window and a maximum of 384K output tokens — three times the 128K ceiling the report cites for the current OpenAI and Anthropic comparison models.

Model identity differs across providers. The first-party endpoint uses the ID deepseek-flash; the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are temporarily accepted and routed to the new release at Flash rates. From September 14, 2026 at 04:00 UTC, DeepSeek says requests for deepseek-v4-pro will also route to V4.1 Flash until a V4.1 Pro model arrives. On the Modelflare gateway the release appears under the versioned ID deepseek-v4.1-flash-0910 with both Chat Completions and Responses interfaces, and the report advises using the exact ID your provider lists rather than assuming a first-party alias carries through every gateway.

An architecture tuned for long-running agents

The model pairs a 552B-parameter mixture-of-experts backbone with a separate 196B-parameter Engram memory store, yet neither figure is per-token compute. About 8B parameters activate for each prompt token during prefill, versus roughly 16B for each decoded token. The 40 transformer layers split into a 20-layer causal encoder and a 20-layer decoder.

That layout, which DeepSeek calls Causal Encoder-Decoder (CED), targets the biggest recurring cost for tool-using agents: reprocessing a history that grows with every observation and tool result. Instead of building an independent global key-value history at every decoder layer, the decoder's global keys and values are projected from the encoder's final hidden state, while local sliding-window states remain layer-specific. DeepSeek's own estimate puts long-sequence prefill work at roughly half of an equivalent all-layer design. The trade-off is that the saving applies to input, not generation — workloads with very large prompts and short answers benefit most.

Shrinking the KV cache

Three further mechanisms attack memory costs. Compressed Sparse Attention 2 (CSA2) statically assigns each layer to one of three modes — Full, Reindex or Reuse — with the decoder's first Full layer selecting a Top-512 candidate pool that later layers draw from. Global KV entries are quantised to roughly four-bit E2M1 values with an E4M3 scale per 16 channels, which DeepSeek reports at 890 bytes per token: around a quarter of V4 Flash and 437 times smaller than V1. The more quantisation-sensitive local sliding-window cache stays in FP8, and a Bounded Replay scheme reconstructs a bounded local window instead of persisting every layer's local state, cutting persistent cache to roughly one-eighth of V4 Flash's footprint. The Engram store moves conditional n-gram patterns into sparsely accessed tables, and DSpark speculative decoding drafts several candidate positions, choosing a verification length based on confidence and system load. Images pass through a DeepSeek-ViT encoder plus a projector, and pretraining covered 45T tokens at a 7:1 text-to-multimodal ratio.

Serving details and limits

Reasoning effort is exposed through three presets mapped to an internal scale: low at 50, high at 75 and max at 100. The official API allows 2,500 concurrent requests. The dev.to report characterises the API as unusually cheap without publishing exact figures, arguing that the model's real edge is how much long-horizon agent work each dollar buys while it stays competitive on coding, terminal, automation and tool-use evaluations.

Caveats

The benchmark figures come from DeepSeek itself and were not independently reproduced by Modelflare. DeepSeek's own technical report concedes the model still trails the largest closed systems on the hardest reasoning and edge cases. The architectural tricks improve serving economics and throughput rather than demonstrating stronger reasoning, and the 384K output ceiling is a limit, not a target: very long generations raise latency, cost and the likelihood of the output drifting from its instructions.

Why it matters

The binding constraint on agentic AI is increasingly the cost of maintaining and reprocessing very long histories. V4.1 Flash attacks that constraint directly — halving prefill work by DeepSeek's estimate, compressing global KV fourfold versus its predecessor, and offering a 1M-token context with 384K output under an MIT licence. Even if it does not top the leaderboards, an open-weight model that makes million-token agent loops economically plausible changes what developers can attempt, and its aggressive pricing puts pressure on every commercial rival serving similar workloads.