NVIDIA published Nemotron-Labs Dory1.0-4B on Hugging Face on Oct 16: a 20-layer shared block runs 1, 2 or 4 times at inference so one checkpoint serves several compute budgets. NVIDIA reports 61.8 on SWE-bench Verified at 4 loops, close to Qwen3.5-9B. BF16/MXFP8/NVFP4/GGUF under OpenMDW-1.1.
Key Takeaways
- ✓Architecture: 8 input + 20 shared recurrent + 8 output layers; loop count 1/2/4 chosen at inference, 2 recommended
- ✓Vendor-reported SWE-bench Verified: 50.3 / 59.5 / 61.8 at 1/2/4 loops vs 61 for Qwen3.5-9B
- ✓Terminal-Bench 2.1 at 4 loops is 22.2, well below Qwen3.6-35B-A3B at 43
- ✓NVFP4 is ~2.17–2.44x faster than BF16 on DGX Spark; last-loop KV caching holds cache at one-loop size
- ✓vLLM (PR #60149) and llama.cpp (PR #30306) support not yet in releases; build from the PRs
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
NVIDIA quietly published Nemotron-Labs Dory1.0-4B on Hugging Face on Oct 16, 2026 (no blog post). It is a looped (latent-recurrent) normalized-GPT: 8 input layers, a 20-layer shared recurrent block that can run 1/2/4 times at inference, and 8 output layers, with interleaved sliding-window and full attention, 256K context, English only. Pre-trained on 6T tokens (natively in NVFP4, then BF16) plus 300B tokens of context extension, post-trained with multi-environment GRPO. A last-loop KV-cache mode keeps cache memory at one-loop size regardless of loop count.
NVIDIA-reported scores (no independent replication yet): SWE-bench Verified 50.3/59.5/61.8 at 1/2/4 loops, on par with Qwen3.5-9B (61) in the same table; Terminal-Bench 2.1 22.2 at 4 loops vs 43 for Qwen3.6-35B-A3B; IFBench 79. NVFP4 runs ~2.17–2.44x faster than BF16 on DGX Spark.
Two loops is the recommended default; four for agentic work. vLLM support is still in review (PR #60149) and llama.cpp support lives in PR #30306, so you build from those branches for now. Weights ship in BF16, MXFP8, NVFP4 and Q5_K_M GGUF under the commercially usable OpenMDW-1.1 license.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.