Agens Volundr 32B v1 Kills Repetition Loops; MIT Engine Fits 262K Context in 16GB
ComfortableKindly507 · reddit · 2026-10-18
Blockway shipped Agens Volundr 32B v1 addressing community complaints from the Preview, and open-sourced the Agens inference engine under MIT.
- Repetition loops on SWE-bench tasks: 36/50 → 0/50; reasoning steps no longer truncated at 32K
- Session continuation makes follow-up turns 2.9x faster (1.6s median) via state reuse instead of re-prefilling
- SWE-bench Verified 50-task subset 22 → 27/50, τ²-bench 89 → 92/120, median SWE task time 124 → 82 min
- Hybrid architecture: 54 of 72 layers use fixed-size state instead of KV cache; decode speed stays flat 1K–128K
- Agens Engine decodes 1.4–1.6x faster than their SGLang build; a 16GB-budget build runs 262K context (validated to 239K) at 2.85 bits/weight, 25GB download
- Honest caveats: GPQA drops to 73.7 vs 80.3 bf16; measured on a 48GB card under a hard 16GiB cap; 151s time-to-first-token on a 219K prompt; LiveCodeBench v6 67.9 and AIME 2025 80.0 beat Qwen3.8-27B
More from Infra
- Agent sandbox showdown: Solo.io compares NVIDIA OpenShell vs kagent + Agent Substrate on scale and HA — bibryam · 2026-10-18
- llama.cpp PR adds IQ2_K-IQ6_K quants: same file size, much better KLD — ethertype · 2026-10-18
- Analyst argues market is badly wrong on Micron: HBM to grow 7x with prices tripling — JOBhakdi · 2026-10-18
- FT: 47 major US AI data center projects delayed or canceled amid $10T buildout — TansuYegen · 2026-10-18
- Google rewrites SRE book for the AI era, announces second edition 10 years later — mgill25 · 2026-10-18
- GPT-2 KV cache from scratch: 1000x fewer FLOPs but only 6x latency gain, measured — atronos_kronios · 2026-10-18