DeepSeek Flash KV Cache at 890 Bytes per Token

890 bytes per token KV cache

The Hugging Face model card states that the Causal Encoder-Decoder design projects the decoder global KV cache from the final encoder hidden states rather than from each decoder layer. Compressed Sparse Attention 2 and FP4 main KV caching (E2M1 format) bring the footprint to 890 bytes per token. Source 1 adds that the new cache requires one-quarter the HBM and one-eighth the SSD storage of the prior generation.

Cache-hit charges form a large share of agent costs on long sessions. The same card notes that SWA Bounded Replay avoids persisting sparse attention states to SSD, further cutting persistent storage. No equivalent KV cache figure appears for GPT-5.6 Sol in any of the four sources. The SCMP report records the Terminal-Bench result but supplies no cache or storage numbers for either model.

DeepSeek-V4.1-Flash activates 8 billion parameters on input and 16 billion on output inside a 552 billion parameter MoE backbone. The card lists a 1 million token context window and a maximum output length of 384 thousand tokens. These specifications sit alongside the KV cache claim and are absent from the GPT-5.6 Sol entries in the same sources. The pre-training corpus reached 45T tokens with sparse attention trained at 64K sequence length before context extension to 1M tokens at 34T tokens.

The technical report further specifies 196B engram conditional memory parameters accessed via token-based lookup and DSpark speculative decoding. Single-Pass mHC revised residual-stream mixing and a vision encoder with 2D-RoPE and 3x3 pixel-unshuffle downsampling complete the multimodal stack. None of these component counts or training token volumes appear for GPT-5.6 Sol in the provided sources.

2500 concurrent requests limit

The API pricing page sets a concurrency limit of 2500 for the deepseek-flash endpoint. The prior V4-Pro endpoint carried a 500 limit. The higher ceiling pairs with the smaller KV cache to keep longer agent sessions inside the same hardware envelope without proportional storage growth.

Off-peak input cache-miss pricing sits at $0.15 per million tokens and peak at $0.30. Output peaks at $1.20. All rates are per million tokens and apply after the September 10 pricing change. The same page states that requests to the retired deepseek-v4-pro endpoint will route to V4.1-Flash at these rates beginning 04:00 UTC on September 14, 2026.

Peak and off-peak windows are defined on the pricing page as 01:00-04:00 and 06:00-10:00 UTC Monday through Friday for peak hours. All other hours count as off-peak and receive half the listed rates. No concurrency or cache-miss pricing appears for GPT-5.6 Sol in the provided sources.

The pricing page also lists support for JSON output, tool calls, Responses API, Anthropic API format, and Chat Prefix Completion. FIM Completion remains limited to non-thinking mode. Thinking mode itself supports both non-thinking and thinking defaults with a continuously controllable reasoning effort setting from 1 to 100. These feature and mode details have no counterpart entries for GPT-5.6 Sol among the sources.

74.2 resolved rate on DeepSWE v1.1

Teams that run multi-turn coding agents or cyber evaluation harnesses over hour-long contexts pay repeated cache-hit fees. The documented one-quarter HBM and one-eighth SSD reduction compounds directly on those sessions. Buyers who already operate within the DeepSeek Harness or OpenCode scaffolds can switch the model string to deepseek-flash and retain the same 1 million token window without code changes.

Teams that require published pricing transparency from every vendor or that depend on undisclosed frontier-model integrations will continue to select GPT-5.6 Sol. The sources give no figure for Sol cache size or concurrency, so direct cost modeling remains impossible on those axes. The architecture favors sustained agent runs where cache costs dominate total spend. Independent harness reproductions have not yet appeared outside the vendor’s own evaluation folder.

WorkBuddy and OpenCode now list full support for the new model. Large-scale deployments with 2,000 GPUs plus a storage cluster receive an open invitation on the announcement page. The model card supplies a self-contained Python reference implementation for prompt encoding that covers tool calling, thinking mode, numeric reasoning effort, and interleaved images, plus Rust libraries with Python bindings for production use.

Until the September 14, 2026 routing change completes and third-party reproductions of the full agent harness results appear, measured KV cache size remains the clearer predictor of sustained coding and cyber workload cost than the two-point Terminal-Bench margin.