Blockway releases Agens Volundr 32B v1 with MIT-licensed inference engine and 262K context support

Blockway has released Agens Volundr 32B v1, addressing critical issues flagged in the previous preview version. The update focuses on stability and efficiency, specifically eliminating repetition loops in agentic tasks and introducing a new open-source inference engine.

Key Improvements and Fixes

The v1 release resolves the primary complaints from the preview phase:

• Loop Elimination: Repetition loops on SWE-bench tasks dropped from 36/50 in the preview to 0/50 in v1.
• Reasoning Efficiency: Median reasoning steps are now 120 tokens (down from cuts at 32K), with the longest step at 12K.
• Session Continuation: The SGLang build now supports session continuation, making follow-up turns 2.9× faster (1.6 s median) by reusing model state instead of re-prefilling.
• Agent Performance: SWE-bench Verified (50-task subset) improved from 22/50 to 27/50, and τ²-bench scores rose from 89/120 to 92/120. Median SWE task time decreased from 124 min to 82 min.

Agens Engine (MIT License)

Blockway is open-sourcing Agens Engine under the MIT license, a custom inference engine designed for the model's hybrid architecture where 54 of 72 layers use fixed-size state instead of a KV cache.

• Performance: The engine decodes 1.4–1.6× faster than Blockway's own SGLang build (up to 1.8× on code with thinking).
• Throughput: Chat speeds increased from 58 to 86 tok/s, code from 79 to 113 tok/s, and JSON generation from 125 to 199 tok/s.
• 16 GB Build: A new Agens-Volundr-32B-v1-Engine-16GB build offers 262K context (measured up to 239K) within a 16 GB memory budget using mixed 2/3/4-bit quantization (~2.85 bits/weight).

Benchmark Results

The model shows competitive performance in coding and math, though some general knowledge scores dipped compared to the preview due to tuning for agent behavior.

• LiveCodeBench v6: 67.9 (best of 2 runs), +8.4 over Qwen3.8-27B.
• AIME 2025: 80.0 (best of 2 runs), +6.7 over Qwen3.8-27B.
• GPQA Diamond: 80.3 (down from 84.3 in preview).
• HumanEval: 78.7 (down from 81.7 in preview).

Limitations and Requirements

• Hardware: The 16 GB engine build requires Linux, NVIDIA driver with CUDA 12.9, and 32 GB+ host RAM for the 262K context window. It was measured on a 48 GB card under a hard 16 GiB cap.
• Compatibility: Stock vLLM cannot load the model. SGLang and llama.cpp require Blockway's specific builds. No Metal/Mac support in the llama.cpp path yet.
• Quality Trade-off: The 16 GB quantized build shows a drop in GPQA (73.7 vs 80.3 at bf16) but maintains coding performance (HumanEval 73.2% vs 73.8% for INT4).