Agentic benchmark results reported by DeepSeek at maximum reasoning effort. See the official model card for evaluation settings and methodology.

DeepSeek-V4.1-Flash is DeepSeek's efficient multimodal model for long-context agents, coding, reasoning, and visual understanding. The 552B-parameter Mixture-of-Experts model activates 8B parameters while processing input and 16B while generating output, with a context window of up to one million tokens. It is available through LM Studio Cloud.

Highlights

• Native text and image input with text output
• Up to a 1M-token context window
• Continuously controllable reasoning effort
• Strong coding, tool-use, and long-horizon agent performance
• MIT-licensed open weights
• Hosted LM Studio Cloud inference without managing a 552B model locally

Asymmetric architecture

V4.1-Flash introduces a Causal Encoder–Decoder architecture designed to spend less compute on long inputs than on generation. Its 552B-parameter backbone activates 8B parameters per token during prefill and 16B during decoding. A native vision encoder processes images alongside text from pre-training rather than adding vision as a separate post-training capability.

DeepSeek combines Compressed Sparse Attention 2 with a hierarchical sparse indexer to reduce long-context inference overhead. The model was trained with context lengths extended to one million tokens and is designed for input-heavy agentic workloads.

Agentic performance

DeepSeek reports the following results at maximum reasoning effort:

Benchmark	DeepSeek V4.1 Flash	Kimi K3	GLM-5.3	Opus 5	GPT-5.6 Sol
Terminal-Bench 3.0	30.0	17.7	28.3	43.3	34.4
DeepSWE v1.1	74.2	67.5	66.9	74.0	73.0
CyberGym	88.1	80.0	84.5	84.5	84.5
AutomationBench	54.8	46.7	48.8	50.3	45.8

V4.1-Flash also reaches 90.6 on Terminal-Bench 2.1, 64.0 on NL2Repo-Bench, and 63.9 on HLE with tools in DeepSeek's reported evaluations. Results depend on the agent scaffold, reasoning effort, sampling settings, and evaluation methodology.

Reasoning controls

DeepSeek-V4.1-Flash supports reasoning effort from 1 to 100. LM Studio presents this as Off, Low, Medium, High, and Max, letting you trade latency and token use for deeper deliberation.

DeepSeek recommends temperature = 1.0, top_p = 0.95 or 1.0, a 1M-token context window, and an output budget of at least 256K tokens for demanding agentic evaluations. LM Studio Cloud supports the 1M context window and a 256K output budget for this model.

Smaller KV cache

DeepSeek reports that V4.1-Flash uses about one quarter of the HBM and one eighth of the SSD storage required by the previous V4-Flash generation's global KV cache. The reduction is designed to improve throughput and lower the cost of cache-heavy, long-running agents.

Use DeepSeek V4.1 Flash in LM Studio

Choose DeepSeek V4.1 Flash from the Cloud model picker in LM Studio or Bionic. It is a strong fit for repository-scale coding, terminal work, research, visual document understanding, and agents that need to retain large working contexts.

License and sources

DeepSeek-V4.1-Flash is released under the MIT License.

DeepSeek announcement · Official model card · Baseten model library · Technical report