Blockway has released Agens Volundr 32B Preview, a dense model of roughly 32 billion parameters across 72 layers, under an Apache-2.0 license. This is not the usual LLM announcement: the story is a hybrid architecture that reduces the KV cache to 18 of 72 layers. For anyone running models on-premise, that choice targets the most concrete constraint in long-context workloads.

In traditional transformers, the KV cache grows with context length and can fill VRAM before the weights do. Blockway starts from field experience: its customers run models on their own machines, so the real limit at long context is memory, not parameter count. The team therefore designed a stack in which most layers do not keep a KV cache.

Of the 72 layers, 54 use Kimi Delta Attention: linear attention with a fixed-size recurrent state and no KV cache. Another 17 use BCSA, a compressed-sparse attention that keeps an exact window over the last 4,096 tokens and pools older context 4:1 into blocks; a learned indexer selects the top 512 blocks. Only layer 72 uses full attention. On top of this, Engram adds hash-based n-gram memory attached to two layers and held in host RAM, while mHC uses four residual streams instead of one. The declared result is a 262K context window.

The speed numbers published by the team show nearly flat decode. On two 48 GB GPUs in BF16, the model produces 25.1 tokens/s at 1K context and 23.9 tokens/s at 128K. Prefill drops from 2,122 to 1,297 tokens/s over the same range. In INT4 at 31.7 GiB on a single 48 GB GPU, decode reaches 31.0 tokens/s at 1K and holds at 29.1 tokens/s at 32K. Aggregate BF16 throughput is 127 tokens/s with 8 users and 130 with 16; INT4 drops to 117 with 8 users. These figures come from the team's own sglang fork, not stock runtimes.

The structural point is not just speed, but where the bottleneck ends up. By shrinking the KV cache and shifting part of the memory to host RAM, this kind of architecture changes hardware sizing for self-hosted deployments. It does not remove the need for GPUs: weights and the exact window still live in VRAM. But it makes it less likely that two or more GPUs are needed only to hold long context. Teams evaluating local infrastructure can look at configurations with a single 48 GB GPU and more system RAM, with direct effects on TCO and operational complexity.

The trade-off shows up in agentic performance. In the published benchmarks, the model beats Qwen3.8-27B on LiveCodeBench v6, HumanEval, AIME 2025, and MATH-500, and is roughly level on MMLU-Pro, IFEval, and GPQA Diamond. It trails on agentic tasks: tau2-bench at 74.2 versus 79-80 from competitors, and SWE-bench Verified at 44 versus 58-64 on a 50-task subset. The team acknowledges this and points to closing the gap as the priority for the full version, with additional pre-training to about 10 billion tokens and training on long agentic sessions.

Then there is the serving question. Today the model runs only with the team's sglang fork: stock sglang and vLLM cannot load it yet, while GGUF and llama.cpp are planned but not available. For real on-premise deployment, this dependence on a custom runtime is the most relevant integration cost. Docker containers are available for 48 GB Ada GPUs and for H100/H200. The DFlash2 drafter, in a separate repository, adds up to 3.6x on JSON and tool output, 2.0x on code, and 1.6x in thinking mode for a single user, but the benefit fades beyond about 8 concurrent users.

Blockway presents this as a preview and asks users to try it, report failures, and suggest benchmarks. The underlying question is not whether the KV cache will disappear, but how quickly standard runtimes can support hybrid architectures like this one. Until they do, the architectural advantage remains tied to the vendor's fork.