The AI Loadout
←Back to models
Checked Active: 2026-10-18
NVIDIA

NVIDIA Dory 1.0 4B (local)

A small reasoning model where you choose how hard it thinks: one to four passes through its shared layers from the same file. For tinkerers comfortable building llama.cpp or vLLM from source.

Evaluation Analysis: Released Oct 16, 2026 by NVIDIA Nemotron Labs. A looped model: a shared 20-layer block runs one to four times per token. NVIDIA recommends two loops as the default and four when accuracy matters more, such as agent tasks. NVIDIA reports SWE-bench Verified at 50.3 with one loop, 59.5 with two and 61.8 with four, against 54.7 for Qwen3.5 4B. Text only, English. Not in Ollama or LM Studio yet: until support lands in releases, build llama.cpp from the Dory pull request (ggml-org/llama.cpp#30306) or install vLLM from vllm-project/vllm#60149, as the model card explains. Weights in BF16, MXFP8, NVFP4 and a Q5_K_M GGUF. Plan on RAM or VRAM of roughly the download size plus 2 to 4 GB for context.
Context WindowUp to 256K
Input Price / 1MFree (local)
Output Price / 1MFree (local)
Release Date2026-10-16

Run it locally

./build/bin/llama-server -m ./dory-q5_k_m.gguf --alias dory --jinja --reasoning on --override-kv "dory.num_loops=int:2,dory.recurrent_kv_cache_mode=str:per_loop"
Ollama tag
Not in the Ollama library yet
LM Studio name
Not in LM Studio yet
Download size by quantization
Q5_K_M GGUF 3.7 GB; BF16, MXFP8 and NVFP4 on Hugging Face
License
OpenMDW 1.1

Memory needed is roughly the download size plus 2 to 4 GB for context. See how to run AI locally.

Specs and prices checked 2026-10-18 against the official page.

Used in these setups

Related skills and personas

Agent skills and bot personas that pair well with NVIDIA Dory 1.0 4B (local).

Browse full directories:All Skills (>35)·All Personas (>15)

Want tested skills and personas installed in one command? See the AI Loadout packs, from $9.