kyojin
Kyojin: the Yamz inference engine for AMD Strix Halo (ROCm, gfx1151), built on ExLlamaV3. Runs 300B-class MoE models on one 128 GB mini PC.
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 226 launches in local AI inference and runtimes — see how it stacks up on momentum and crowding →
69 other launches read as similar to this one →
Details
- External ID
- 1402338989
- Source
- GITHUB
- Company
- —
- Product
- kyojin
- Website domain
- huggingface.co
- Launched
- Oct. 2, 2026
- Cohort
- —
- Upvotes
- 53
- Upvotes percentile
- 0.7555960437272254
- Tags
- amd, exl3, inference, llm, quantization, rocm, strix-halo
- Fetched at
- Oct. 6, 2026, 5:02 p.m.
- Updated at
- Oct. 6, 2026, 5:02 p.m.
Enrichment
- Niche
- local AI inference and runtimes
- Vertical
- Horizontal
- Function
- Model & infra
- Audience
- Developer
- AI stance
- AI-native
- Project type
- Hobby / open-source project
- Normalized one-liner
- inference engine for running large moe models on amd strix halo
- Manually corrected
- False
Could you build this?
No Kyojin is a low-level GPU inference engine optimizing ExLlamaV3 kernels specifically for AMD ROCm and RDNA 3.5 architecture (gfx1151).
What it would actually take: Building this engine requires deep knowledge of AMD HIP/ROCm, gfx1151 ISA, and custom GPU compute kernel development (assembly or low-level HIP C++). The core challenge is writing bespoke quantized matrix multiplication (GEMM/GEMV) kernels and memory bandwidth schedulers to maximize APU unified memory throughput for 300B MoE models. Requires senior systems software engineers specializing in compiler internals and GPU microarchitecture.
Competitors
Other products that read as similar to this one — 69 launches clear the similarity bar, closest 8 shown.
Attention rank: #17 of 70 (itself plus its competitors, highest first — normalized so YC and Product Hunt are compared fairly).
Launched 328 days after the earliest competitor.
- whirl-llm · github · 2026-10-03 · 15 upvotes · similarity 0.41
- Open-source AMDGCN kernels for optimizing LLM inference · hn · 2026-08-25 · 5 upvotes · similarity 0.41
- Goku · hn · 2026-07-15 · 9 upvotes · similarity 0.40
- A new engine to run Kimi K3 on a laptop · hn · 2026-07-29 · 7 upvotes · similarity 0.39
- exl3xpu · github · 2026-09-22 · 10 upvotes · similarity 0.39
- GSQHalo.cpp · github · 2026-10-02 · 9 upvotes · similarity 0.39
- Janus · hn · 2026-10-01 · 106 upvotes · similarity 0.39
- MegaCapybara · github · 2026-10-03 · 26 upvotes · similarity 0.37
Other launches for this product
- No other launches for this product.
Same idea, different domain
Nobody's really built a model & infra tool for Fintech yet.