DeepSeek-V4.1-Flash

How it is built

DeepSeek-V4.1-Flash splits its 40 layers in two: a causal encoder of 20 layers, then a decoder of 20.¹¹1DeepSeek, DeepSeek-V4.1-Flash model card, 10 September 2026. Instruct benchmarks at reasoning_effort=100, temperature 1.0, top_p 0.95; Terminal-Bench 2.1 Pass@1 in the Minimal mode of DeepSeek Harness with a 1M-token context. Base models on 25-shot SimpleQA-Verified (exact match): V4.1-Flash 42.3, V4-Pro 55.2. https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash The decoder does not build its global KV cache from its own layers. It projects the cache from the encoder's final hidden states.¹¹ DeepSeek gives this as the reason a token needs only 8B parameters in prefill and 16B in decode.¹¹

The backbone counts 552B parameters.¹¹ The Engram memory is listed apart at 196B. The card does not say whether the 552B includes it.¹¹

In each MoE layer a token reaches 6 of 384 routed experts, plus 1 shared expert.²²2DeepSeek, DeepSeek-V4.1-Flash config.json (n_routed_experts 384, n_shared_experts 1, num_experts_per_tok 6, num_hidden_layers 40, sliding_window 128, fp8 weights in 32 x 32 blocks with fp4 experts, vision encoder of 32 layers). https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/main/config.json Attention is Compressed Sparse Attention 2 (CSA2). Each attention layer is fixed to one of three modes, and through them layers share their KV cache and reuse the tokens that an earlier layer's sparse attention picked.¹¹ A 128-token sliding window runs beside it.²² Images go through a vision encoder of 32 layers and enter the model with the text.¹¹²²

What is new

Weights and API came out together on 10 September 2026.³³3DeepSeek API Docs, DeepSeek-V4.1-Flash release note, 10 September 2026. https://api-docs.deepseek.com/news/news260910

The main KV cache is now stored in FP4. At 890 bytes per token the global cache is about a quarter the size of DeepSeek-V4-Flash's.¹¹ A single sequence of 1,048,576 tokens fills about 933 MB.⁴⁴4Arithmetic on the model card figure: 890 bytes x 1,048,576 tokens = 933,232,640 bytes, about 933 MB. https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash What stays on SSD shrinks to about an eighth of V4-Flash's.¹¹

Training started from scratch on a multimodal corpus of 45T tokens. The context was stretched to 1M tokens at the 34T mark.¹¹ Reasoning effort takes any integer from 1 to 100.¹¹

DeepSeek's own table puts V4.1-Flash ahead of the larger DeepSeek-V4-Pro on Terminal-Bench 2.1, at 90.6 against 87.9.¹¹ On factual recall its base model trails.¹¹

Running it

The weights are MIT-licensed. They are stored in FP8 in blocks of 32 by 32, with the routed experts in FP4.¹¹²² No Jinja chat template comes with them, and DeepSeek ships a Python reference encoder in its place.¹¹ The card asks for a 1M context window and a max_tokens of at least 256K.¹¹

The API calls the model deepseek-flash, with 1M tokens of context and up to 384K of output.⁵⁵5DeepSeek API Docs, Models and Pricing, as listed on 29 September 2026. Off-peak covers all other hours, including weekends and Chinese public holidays. https://api-docs.deepseek.com/quick_start/pricing The price depends on the hour. From 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, a million input tokens cost $0.30, or $0.006 from the cache, and a million output tokens $1.20.⁵⁵ Every other hour costs half.⁵⁵