DeepSeek-V4.1-Flash

• Finetunes: 3 models
• Quantizations: 23 models (llama.cpp, LM Studio, Jan, Ollama targets)
• Spaces: 4 spaces using this model
• Collection: part of the DeepSeek-V4 collection (10 items)
• Base model for community GGUF variants (e.g. FP8/8-bit/NVFP4 builds)

@misc{deepseekai2026deepseekv41flash,
      title={DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression},
      author={DeepSeek-AI},
      year={2026},
}

The repository and the model weights are licensed under the MIT License (see the LICENSE file in the repository).

Code agent benchmarks (Terminal-Bench 2.1/3.0/4.0, DeepSWE v1.1, NL2Repo-Bench, ProgramBench) are evaluated with the Minimal mode of DeepSeek Harness and a 1M-token context window; the mini-SWE harness is used for DeepSWE v1.1 and the Claude Code harness for SEC-Bench Pro. Visual agent benchmarks use the Claude Code harness with 512K context. Scaffold comparison (max effort):

All scaffolds: N=8 samples (DeepSWE) / N=3 (Terminal-Bench 2.1), temperature=1.0, top_p=0.95, 1M-token context limit, max_steps=500, no network access for Terminal-Bench 2.1. The evaluation/ folder reproduces the DeepSWE v1.1 results.

This release does not include a Jinja-format chat template. The encoding/ folder contains a self-contained Python reference implementation (encoding.py) with test cases for multi-turn conversations, tool calling, thinking mode, numeric reasoning effort, mid-conversation system messages, and interleaved image content.

For production use, deepseek-recipe (github.com/deepseek-ai/deepseek-recipe) provides the same prompt format as maintained Rust libraries with Python bindings, covering Messages, Chat Completions, and Responses API requests. The inference/ folder documents weight conversion and local inference.

Evaluations use temperature=1.0, top_p=0.95.

The model supports a continuously controllable reasoning effort setting - an integer from 1 to 100 - that trades inference cost for accuracy. All published instruct results use the maximum effort setting (reasoning_effort=100).

The post-training recipe follows the standard SFT -> RL -> on-policy distillation (OPD) paradigm without algorithmic modifications. All substantive changes lie in the data pipeline: large-scale automated synthesis of agent tasks and environments with progressive scaling of data, tasks, and rollouts.

DeepSeek-V4.1-Flash is trained from scratch on a multimodal corpus comprising 45T tokens. Sparse attention is trained at a sequence length of 64K, and context is extended to 1M tokens at the 34T-token mark.

A vision encoder (DeepSeek-ViT, trained from scratch with 2D-RoPE and 3x3 pixel-unshuffle downsampling) and a two-layer MLP projector convert images into visual embeddings, which are processed jointly with text embeddings from the start of language-model pre-training. Base-model multimodal results: MMMU-Pro 56.5, CVBench 77.9, DocVQA 95.6, RefCOCO-avg 86.0.

• Sparse MoE: 1 shared expert + 384 routed experts per MoE layer, activating 6 routed experts per token
• Engram conditional memory: 196B parameters, sparsely accessed via token-based lookup
• DSPark speculative decoding: semi-autoregressive draft generation with confidence-scheduled verification
• Single-Pass mHC: revised residual-stream mixing with an efficient Mega-mHC kernel

Compressed Sparse Attention 2 (CSA2) assigns each attention layer one of three static modes - Full, Reindex, or Reuse - to share main KV and indexer K across layers and reuse Top-K sparse-attention indices. In the decoder, a Hierarchical Sparse Indexer restricts later indexing layers to a candidate pool constructed by the first Full Mode layer, bounding deeper indexer cost independently of context length.

Combined with FP4 main KV caching (E2M1 format, one E4M3 scale per 16 channels), the global KV cache footprint drops to 890 bytes per token - roughly 1/4 of DeepSeek-V4-Flash. SWA Bounded Replay reconstructs missing sliding-window-attention KV states by replaying only the most recent n_win tokens, avoiding the need to persist SWA KV to SSD and reducing the persistent KV cache footprint to about 1/8 of DeepSeek-V4-Flash.

DeepSeek-V4.1-Flash organizes its 40 Transformer layers as a 20-layer causal encoder followed by a 20-layer decoder. With CED, the decoder's global KV cache is projected from the final encoder hidden states rather than derived from each decoder layer's own hidden states. This allows the model to activate only 8B parameters per token during prefill and 16B during decode, substantially improving cost efficiency for input-heavy agentic workloads.

• Multimodal MoE with 552B backbone parameters (763B total), MIT-licensed
• 1M-token context window, native image + text input, autoregressive text output
• 8B parameters active per token during prefill, 16B during decode
• Global KV cache: 890 bytes per token (~1/4 of DeepSeek-V4-Flash, ~1/437 of DeepSeek-V1)
• Records on DeepSWE v1.1 (74.2) and Terminal-Bench 2.1 (90.6) at max reasoning effort
• Continuously controllable reasoning effort (integer 1-100)