ds4
Local inference engine for DeepSeek V4, GLM, Qwen
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 131 launches in open-source LLM deployment and inference — see how it stacks up on momentum and crowding →
179 other launches read as similar to this one →
Details
- External ID
- 1268037
- Source
- PH
- Company
- —
- Product
- dcs4
- Website domain
- producthunt.com
- Launched
- Oct. 7, 2026
- Cohort
- —
- Upvotes
- 1
- Upvotes percentile
- 0.33
- Tags
- Open Source, Developer Tools, Artificial Intelligence
Description
DwarfStar 4 (ds4) is a narrow C inference engine from Salvatore Sanfilippo (antirez) for running DeepSeek V4/V4.1 Flash, GLM 5.x and Qwen3.8 Flash Next locally on high-memory Mac, CUDA and ROCm machines. It uses asymmetric quantization on routed experts and a KV cache that persists to SSD and survives restarts. Ships a CLI, OpenAI/Anthropic-compatible local APIs, and a native coding agent sharing the same model state. MIT licensed. Benchmark: 34.4 t/s generation on M5 Max 128GB.
Enrichment
- Niche
- open-source LLM deployment and inference
- Vertical
- Horizontal
- Function
- Model & infra
- Audience
- Developer
- AI stance
- AI-native
- Project type
- Hobby / open-source project
- Normalized one-liner
- local llm inference engine for developers
- Manually corrected
- False
Could you build this?
No Writing a custom low-level C inference engine with asymmetric quantization, custom GPU kernels (CUDA/ROCm/Metal), and SSD-persisted KV caching requires deep systems programming and machine learning systems expertise.
What it would actually take: Requires deep C/C++ systems engineering, low-level GPU programming (CUDA, ROCm, Metal Shading Language), and advanced knowledge of LLM architectures and quantization math. The system must manage asymmetric quantization kernels, memory paging, asynchronous SSD I/O for KV caching, and multi-GPU tensor distribution. A standard vibe-coding workflow cannot optimize custom matrix multiplication and attention kernels to compete with or build such engines.
Competitors
Other products that read as similar to this one — 179 launches clear the similarity bar, closest 8 shown.
Attention rank: #119 of 180 (itself plus its competitors, highest first — normalized so YC and Product Hunt are compared fairly).
Launched 333 days after the earliest competitor.
- strixite · github · 2026-10-04 · 11 upvotes · similarity 0.48
- dsv41-flash-pp4-170hx · github · 2026-09-14 · 7 upvotes · similarity 0.46
- qwen38-flash-next-nvidia-nvfp4-sm121-sglang · github · 2026-09-12 · 9 upvotes · similarity 0.44
- DeepSeek-V4 · ph · 2026-04-24 · 393 upvotes · similarity 0.44
- Warp · hn · 2026-09-15 · 13 upvotes · similarity 0.43
- Infernix · github · 2026-10-08 · 98 upvotes · similarity 0.43
- Fastest Qwen 3.8 27 on single RTX5090 · ph · 2026-09-07 · 1 upvotes · similarity 0.43
- TandemLLM · github · 2026-09-28 · 21 upvotes · similarity 0.42