nanoswe/nanoswe-192h-261006
archived · low-signalmodel
A 1.7B-parameter SWE agent model trained from random init in 192 B200-hours (one 8×B200 node, 24.0 h of training wall-clock) in two stages: a short web-text pretraining stage (6.7B tokens of FineWeb-Edu), then 37.8B tokens of multi-turn tool-calling SWE agent trajectories — no instruction tuning… Dense, 2.8B params. Context: 32K tokens (config). License: mit. Vendor-reported: "It resolves 25.47% of SWE-bench Verified (pass@1 over K=5 samples on the 483-instance cluster-runnable subset), the new record for the 192 B200-hour budget."
- Vendor-reported: It resolves 25.47% of SWE-bench Verified (pass@1 over K=5 samples on the 483-instance cluster-runnable subset), the new record for the 192 B200-hour budget.
- Org
- nanoswe
- Released
- 2026-10-06
- Status
- archived · low-signal
- License
- mit (permissive)
- Parameters
- 2.8B
- Architecture
- dense · NanoChatForCausalLM
- Context
- 32,768 tokens
- Task
- text-generation · text
- Signal
- 7 HF likes · 717 downloads
- Archived
- 2026-10-19 (no longer listed by the collector; stats as last seen)
Runs on
mixed: real files + estimate; need = weights + FP16 KV cache for 8K tokens + 1GB. A sizing calculation, not tested.
- Q4 ~4.8 GB (estimated size): no listed device
- Q8 ~6.1 GB (estimated size): 8GB GPU (4060 / 3070)
- FP16 ~8.7 GB: 12GB GPU (RTX 3060), 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), 24GB GPU (3090 / 4090), 32GB GPU (5090), 48GB GPU (RTX 6000 Ada / A6000), 96GB GPU (RTX Pro 6000 Blackwell), 1× DGX Spark (128GB), 2× DGX Spark (256GB), 16GB GPU (7800 XT / 9070 XT), 24GB GPU (7900 XTX), Strix Halo 128GB unified (~96GB usable), Arc B60 24GB, Arc B70 32GB, Mac 16GB, Mac 24GB, Mac 36GB, Mac 48GB, Mac 64GB (M1 Max), Mac 96GB, Mac 128GB, Mac 192GB, Mac 256GB, Mac 512GB, CPU only, 32GB RAM (slow), CPU only, 64GB RAM (slow), CPU only, 128GB RAM (slow), CPU only, 256GB RAM (slow)
Links
First published here 2026-10-06 · updated 2026-10-19 · markdown · JSON