DeepSeek-V4.1-Flash-1017

About

DeepSeek-V4.1-Flash-1017 is the October 17, 2026 checkpoint of DeepSeek's MIT-licensed DeepSeek-V4.1-Flash — a minor update that brings substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks while reducing task execution time and output length, advancing the performance–cost Pareto frontier across multiple agent harnesses. Architecture and parameter counts are unchanged: it is a 763B-parameter Mixture-of-Experts (MoE) transformer (as with the base release, roughly 8B parameters are activated per token during prefill and 16B during decode), natively processes images and text (DeepSeek-ViT vision encoder plus a two-layer MLP projector) and generates text, with English and Chinese as primary languages and a 1M-token context window.

Shared V4.1-Flash platform design:

• Causal Encoder-Decoder (CED): a 40-layer Transformer split into a 20-layer causal encoder and a 20-layer decoder; the decoder's global KV cache is projected from the final encoder hidden states rather than from each decoder layer's own states.

• Compressed Sparse Attention 2 (CSA2): every attention layer runs in one of three static modes (Full, Reindex, Reuse), sharing main KV and indexer K across layers and reusing Top-K sparse-attention (DSA) indices; a Hierarchical Sparse Indexer with Top-K of 512 plus a 128-token sliding window bounds indexer cost independently of context length.

• FP4 main KV caching and SWA Bounded Replay, which reconstructs missing sliding-window-attention KV states by replaying only recent tokens.

• Sparse MoE: 1 shared expert plus 384 routed experts per layer with 6 routed experts activated per token (sqrtsoftplus scoring, noaux_tc routing, routed scaling factor 1.5); Engram conditional memory for token-based n-gram lookup; DSpark speculative decoding; and 3 Multi-Token Prediction (MTP) layers.

Release specifics:

• Prompt encoding: no Jinja-format chat template; a self-contained encoding.py reference covers multi-turn conversations, tool calling, thinking mode, numeric reasoning effort, mid-conversation system messages, and interleaved image content. For production use, the deepseek-recipe Rust toolkit converts Messages, Chat Completions, and Responses API requests into DeepSeek V4/V4.1 prompts or token IDs and parses complete or streamed outputs (thinking, tool calls, images, generation settings).

• Recommended sampling: temperature 1.0, top_p 1.0, a 1M-token context window, max_tokens ≥ 256K.

Compared with the September DeepSeek-V4.1-Flash release, this checkpoint reports: Terminal-Bench 4.0 (Claude Code harness) 42.3 vs 31.2, DeepSWE v1.1 (Mini-SWE-V2 harness) 77.0 vs 74.2, ProgramBench 29.1 vs 20.3, NL2Repo-Bench 68.2 vs 65.4, Automation-Bench v1.0.6 60.5 vs 54.8, Agents' Last Exam 35.4 vs 31.8, Apex Bench v1.1 72.1 vs 64.6, and tool-assisted vision gains — Chartography 83.0, BabyVision 92.5, ZeroBench-main 51.0 — plus HLE w/o tools 42.6 (44.0 on the text-only subset). The underlying technical report is "DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression"; the model is maintained by DeepSeek-AI under the MIT License.