Agens Volundr 32B Preview
Open 32B model: only 18 of its 72 layers keep a KV cache
Get picks like this daily. The day's top launches, AI/tech news, and a weekly opportunity spotlight — straight to your inbox.
This is 1 of 143 launches in large language models — see how it stacks up on momentum and crowding →
24 other launches read as similar to this one →
Details
- External ID
- 1269777
- Source
- PH
- Company
- —
- Product
- Agens Volundr 32B Preview
- Website domain
- producthunt.com
- Launched
- Oct. 5, 2026
- Cohort
- —
- Upvotes
- 1
- Upvotes percentile
- 0.33
- Tags
- Open Source, Developer Tools, Artificial Intelligence
Description
Agens Volundr is a dense 32B open-weights model for agents and long working sessions. Only 18 of its 72 layers keep a KV cache; the other 54 are linear attention with a fixed-size state, so long context stays affordable on hardware you own. Decode holds at 23.9 tok/s even at 128K context (BF16, two 48 GB GPUs, single user). 262K window, Apache-2.0, BF16 and INT4 (one 48 GB GPU). Runs on our open sglang build; GGUF planned.
Enrichment
- Niche
- large language models
- Vertical
- Horizontal
- Function
- Model & infra
- Audience
- Developer
- AI stance
- AI-native
- Project type
- Hobby / open-source project
- Normalized one-liner
- open 32b llm with reduced kv cache for developers
- Manually corrected
- False
Could you build this?
No Training or fine-tuning a novel 32B hybrid linear/KV-attention architecture requires deep machine learning research, custom CUDA kernels, and hundreds of thousands of dollars in high-end GPU compute.
What it would actually take: Developing this requires ML researchers designing custom hybrid architectures (e.g., combining Mamba/linear attention with multi-head attention), writing custom Triton/CUDA kernels for hybrid state inference, and pretraining/continual pretraining across large token datasets on a multi-node H100 cluster.
Competitors
Other products that read as similar to this one — 24 launches clear the similarity bar, closest 8 shown.
Attention rank: #16 of 25 (itself plus its competitors, highest first — normalized so YC and Product Hunt are compared fairly).
Launched 223 days after the earliest competitor.
- Cuts Long Horizon Inference Costs by 50% via external KV Cache Offload · hn · 2026-07-26 · 22 upvotes · similarity 0.43
- RunNburn · hn · 2026-07-30 · 11 upvotes · similarity 0.40
- ExANS · hn · 2026-08-05 · 15 upvotes · similarity 0.38
- Mistral Large 4 · ph · 2026-10-07 · 183 upvotes · similarity 0.34
- Taliesin · hn · 2026-06-04 · 10 upvotes · similarity 0.33
- Ontological Directed Synthesis Network · ph · 2026-09-22 · 1 upvotes · similarity 0.33
- Moonshine Open-Weights STT models · hn · 2026-02-24 · 316 upvotes · similarity 0.33
- agent-gpu-calculator · github · 2026-09-24 · 7 upvotes · similarity 0.33
Other launches for this product
- No other launches for this product.