Nicheloom visitors: 14,531 all-time · 8,930 in the last 7 days

The opportunity tracker for new startups.

Today: 1,383 launchs · Niche of the day: document and PDF processing tools · See full report →

cyber-on-strata

在 8 GB 显存的显卡上跑 35B MoE(Cyber-Tiel-Coder-35B-A3B):专家权重流式上卡、未命中专家由 CPU 池就地计算。Strata 引擎的 fork(MIT)。RTX 4080S 实测 8 GB 档 57.7–63.5 tok/s、全常驻 134–136 tok/s。Build from THIS repo — upstream does not support this model.

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 94 launches in llm inference engines — see how it stacks up on momentum and crowding →

158 other launches read as similar to this one →

Details

External ID
1411842428
Source
GITHUB
Company
—
Product
cyber-on-strata
Website domain
huggingface.co
Launched
Oct. 9, 2026
Cohort
—
Upvotes
17
Upvotes percentile
0.36
Tags
cuda, expert-offload, gguf, llm-inference, low-vram, moe, qwen35moe, speculative-decoding

Enrichment

Niche
llm inference engines
Vertical
Horizontal
Function
Model & infra
Audience
Developer
AI stance
AI-native
Project type
Hobby / open-source project
Normalized one-liner
low-vram moe inference engine for local gpus
Manually corrected
False

Could you build this?

No This is a custom low-level C++/CUDA inference engine fork implementing dynamic weight streaming and CPU-GPU heterogenous expert dispatch for 35B MoE architectures on constrained VRAM.

What it would actually take: Requires deep C++/CUDA GPU systems engineering, memory paging orchestration over PCIe, custom matrix kernel optimization, and intimate familiarity with llama.cpp/Strata architectures. Developers must manage asynchronous PCIe transfers and custom MoE routing kernels to execute missing experts on CPU threads without stalling GPU pipelines.

Competitors

Other products that read as similar to this one — 158 launches clear the similarity bar, closest 8 shown.

Attention rank: #105 of 159 (itself plus its competitors, highest first — normalized so YC and Product Hunt are compared fairly).

Launched 342 days after the earliest competitor.

Other launches for this product