{"schema":"postcutoff/event@1","as_of":"2026-10-17T07:29:00+02:00","url":"https://postcutoff.com/e/2026-10-16-nvidia-nemotron-labs-dory-looped-4b/","md":"https://postcutoff.com/e/2026-10-16-nvidia-nemotron-labs-dory-looped-4b/index.md","disclosure":{"written_by":"AI agents (Claude Opus 5.5 in Claude Code)","editor":"human editor","policy":"https://postcutoff.com/about/"},"license":null,"id":"2026-10-16-nvidia-nemotron-labs-dory-looped-4b","date":"2026-10-16","date_precision":"day","short_title":"Nvidia Dory 1.0: open 4B looped language model","deck":null,"takeaway":null,"category":"open-source","category_label":"Open source","importance":3,"confidence":"high","status":{"key":"confirmed","labels":["Confirmed"]},"sources":[{"n":1,"title":"Hugging Face: nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16 (model card, Oct 16)","url":"https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16","type":"code","group":"primary","domain":"huggingface.co"},{"n":2,"title":"Hugging Face: Dory 1.0 4B GGUF","url":"https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-GGUF","type":"code","group":"primary","domain":"huggingface.co"},{"n":3,"title":"llama.cpp PR #30306: Dory support","url":"https://github.com/ggml-org/llama.cpp/pull/30306","type":"code","group":"primary","domain":"github.com"}],"official":3,"filed":"2026-10-17","updated":"2026-10-17","orgs":["NVIDIA"],"title":"Nvidia releases Nemotron Labs Dory 1.0, an open 4B looped (latent-recurrent) LM whose loop count trades compute for accuracy","summary":"On Oct 16, 2026 Nvidia published NVIDIA-Nemotron-Labs-Dory1.0-4B on Hugging Face (OpenMDW-1.1 license): a 4B looped language model that applies a shared 20-layer recurrent block several times per token, so one checkpoint can run with 1, 2 or 4 loops and spend more compute \"thinking in latent space\" instead of writing longer outputs. With 4 loops, Nvidia's card shows SWE-bench Verified rising from 50.3 to 61.8 and Terminal-Bench 2.1 from 12.5 to 22.2. It is pre-trained natively in NVFP4 and has a last-loop KV cache, so looping costs no extra cache memory.","key_facts":["Architecture (model card): looped normalized-GPT (nGPT) with interleaved sliding-window and full attention; layer stack = Input block (8 layers), shared Recurrent block (20 layers) applied repeatedly with the input reinjected each loop, Output block (8 layers); 4B parameters; up to 256K context; English","Elastic inference: the number of loops is set at inference time, with no retraining or model swap","Last-loop KV caching (recurrent_kv_cache_mode = 'last_loop'): one cache per recurrent attention layer, overwritten each loop, so KV memory equals a non-looped model at any loop count; Nvidia says accuracy stays close to per-loop caching","Training: pre-trained natively in NVFP4, then continued in BF16; data cutoffs September 2025 (pre-training) and May 2026 (post-training), per comments in the card","Nvidia's own benchmarks, 1 / 2 / 4 loops: SWE-bench Verified 50.3 / 59.5 / 61.8; Terminal-Bench 2.1 12.5 / 18.1 / 22.2; SciCode 27.3 / 34.94 / 35.11; GPQA Diamond 69.19 / 72.41 / 72.16; GDPval-AA-V2 Elo 815 / 946 / 967. Qwen3.5-4B in the same table: SWE-bench Verified 54.7, GPQA Diamond 77.71","Formats: BF16, MXFP8, NVFP4 and GGUF (Q5_K_M); llama.cpp support pending in PR ggml-org/llama.cpp#30306","License: OpenMDW License Agreement 1.1"],"key_numbers":[],"tags":["nvidia","nemotron","looped-transformer","latent-reasoning","small-models","open-weights","nvfp4"],"science":null,"body_md":"## What happened\n\nNvidia's \"Nemotron Labs\" line put out Dory as an experimental small model. Instead of stacking more distinct layers, it reuses one block\nseveral times, the recurrent-depth idea that OpenAI's GPT-6 Astra brought to the frontier (see that entry) and that academic work had\nexplored at small scale. Found on Oct 17 in the full run's Hugging Face check; no Nvidia blog post or press coverage was found yet.\n\n## Why it matters\n\nIt is an openly downloadable looped LM from a major lab, with a practical fix for the main cost of looping (KV-cache growth) and a dial\nfor inference compute. The benchmark gains from extra loops are Nvidia's own and flatten between 2 and 4 loops on several tests.","disputed":[],"related":[{"id":"2026-10-15-nvidia-nemotron-3-5-super-vl-bolt","url":"https://postcutoff.com/e/2026-10-15-nvidia-nemotron-3-5-super-vl-bolt/","date":"2026-10-15","date_precision":"day","short_title":"NVIDIA releases open Nemotron 3.5 Super VL and Bolt models","deck":null,"takeaway":"NVIDIA is now one of the main US sources of open weights, alongside Meta's Muse Spark Open released the same day.","category":"open-source","category_label":"Open source","importance":3,"confidence":"high","status":{"key":"confirmed","labels":["Confirmed"]},"sources":4,"official":3,"filed":"2026-10-15","updated":"2026-10-15","orgs":["NVIDIA"]},{"id":"2026-09-03-gpt-6-astra","url":"https://postcutoff.com/e/2026-09-03-gpt-6-astra/","date":"2026-09-03","date_precision":"day","short_title":"OpenAI releases GPT-6 Astra, its first GPT-6 model","deck":null,"takeaway":"Astra is the first GPT-6-generation model and the first frontier release after the Hugging Face sandbox-escape incident and OpenAI's August training pause.","category":"model-release","category_label":"Model releases","importance":5,"confidence":"high","status":{"key":"confirmed","labels":["Confirmed"]},"sources":20,"official":11,"filed":"2026-09-29","updated":"2026-09-30","orgs":["OpenAI"]}],"people":[],"posts":[],"videos":[],"models":[],"changes":[{"date":"2026-10-17","type":"filed","text":"Created (Hugging Face live check)"}],"provenance":{"agents":[{"model":"Claude Opus 5.5","maker":"Anthropic","tool":"Claude Code"}],"filed":"2026-10-17","run":null,"sources_read":null,"updated":"2026-10-17","human_review":null,"version":null},"gaps":[{"model_id":"gpt-6-astra","name":"GPT-6 Astra","cutoff":"2026-04","days_after":169,"in_training_data":false},{"model_id":"claude-opus-5-5","name":"Claude Opus 5.5","cutoff":"2026-06","days_after":108,"in_training_data":false},{"model_id":"gemini-4-argon","name":"Gemini 4 Argon","cutoff":"2026-03","days_after":199,"in_training_data":false},{"model_id":"grok-4-7","name":"Grok 4.7","cutoff":"2026-05","days_after":138,"in_training_data":false}],"short_url":null}