llama.cpp release b11870 (25 Oct) added support for DeepSeek V4.1. The 17 Oct build has the joint-highest score among downloadable models on the Artificial Analysis Intelligence Index v5 (32), with vision, a 1M-token context and an MIT licence.
One of the two highest-scoring open models no longer needs special builds to run locally. Running it in-house sends no prompts or data to the lab, but check any US federal contract terms first.
What happened
DeepSeek released V4.1-Flash-1017 on 17 October 2026. On 25 October, llama.cpp release b11870 added support for DeepSeek V4.1, so the model now runs in mainline llama.cpp, the widely used open-source runtime for running models on your own hardware.
Key facts
- Lab: DeepSeek (China)
- Released: 17 Oct 2026
- Parameters: The learned numbers inside a model. More parameters usually means a more capable but larger model. Glossary: 552B total (763B including Engram), Mixture of experts (active parameters): A design where only part of the model (the active parameters) works on each piece of text. Glossary
- Artificial Analysis Index v5: 32, the joint-highest score among released open weights, with GLM-5.3
- Features: vision, Context window: How much text a model can take into account at once, measured in tokens. Glossary
- Licence: MIT
- 4-bit download (Unsloth UD-Q4_K_XL): 500.7 GB
Our view
Mainline support matters because it removes the need for patched or third-party builds, which are harder to maintain and to trust. As a rule of thumb, a model of about 500 GB needs roughly that much combined RAM and VRAM to run at usable speed, so this is a server-class deployment.
Running the weights locally sends no prompts or data to DeepSeek. However, the US has ordered DeepSeek off federal civilian IT systems (see our item on the DHS order), so organisations with US federal contracts should check their obligations first.
Why it matters: Running it in-house keeps prompts and data on your own equipment, but it needs server-class hardware and a check of any US federal contract terms.
Source: llama.cpp (ggml-org, GitHub), “Release b11870”, 25 Oct 2026