lightweight and on-device AI runtimes
Recent window: last 5.1 months
(2026-04-24 → 2026-09-27), compared with the prior
5.1 months.
horizontal · 264 members ·
Data as of 2026-09-30
These products provide compact models, distillation techniques, and optimized runtime engines designed to run AI locally on consumer hardware and edge devices. They are built for developers, hardware hackers, and embedded systems engineers seeking private, low-latency machine learning execution without expensive cloud GPUs. Unlike general cloud-hosted LLM APIs, this cluster focuses strictly on extreme memory efficiency and resource-constrained local inference.
Metrics
- Stage
- crowded
- Recent count
- 134
- Prior count
- 109
- Total count
- 264
- Momentum
- 22.94
- Attention
- 0.54
- Crowding
- 0.73
- Concentration
- 0.83
- Opportunity
- 0.32
Opportunity components
- Attention
- 0.54
- Low crowding
- 0.27
- Momentum (normalized)
- 0.31
- Low concentration
- 0.17
Source split
- github
- 10 (0.07)
- hn
- 91 (0.68)
- ph
- 33 (0.25)
- yc
- 0 (0.00)
Dominant source: hn · Divergence: 0.68
Members
|
Name
|
Source
|
Upvotes ▲
|
Launched
|
| Smaller than WinRaR but 4x faster |
PH |
1 |
2026-09-18 |
| Natyv |
PH |
1 |
2026-09-16 |
| Foretop |
PH |
1 |
2026-09-18 |
| The Type 1 Civilization Toolkit |
PH |
1 |
2026-09-24 |
| Windows 11 ARM64 on M1 Mac, not in a VM |
PH |
1 |
2026-09-18 |
| HUPI |
PH |
1 |
2026-09-18 |
| VRAMGlass |
PH |
1 |
2026-09-19 |
| Maliklang-V4 |
PH |
1 |
2026-09-30 |
| TritonX |
PH |
1 |
2026-09-06 |
| GOSH.AI DePools |
PH |
1 |
2026-09-25 |
| Tom |
PH |
1 |
2026-09-18 |
| Fastest Qwen 3.8 27 on single RTX5090 |
PH |
1 |
2026-09-07 |
| lepoch |
PH |
1 |
2026-09-16 |
| Aerion |
PH |
1 |
2026-09-13 |
| Efficio - AI Harness for speed, memory |
PH |
2 |
2026-09-10 |
| BestLLMfor |
PH |
2 |
2026-09-09 |
| UsingOpen |
PH |
2 |
2026-09-26 |
| ElideDB. Database for Physical AI |
PH |
2 |
2026-09-18 |
| Autotune Doctor |
PH |
2 |
2026-09-20 |
| Lifeboat |
PH |
2 |
2026-09-23 |
| ROCmFix & InferBench |
PH |
2 |
2026-09-20 |
| TRLoom |
PH |
3 |
2026-09-15 |
| builtwithlaya |
PH |
3 |
2026-09-24 |
| Mercury 2.5 |
PH |
3 |
2026-09-09 |
| Rust-split |
HN |
5 |
2026-09-12 |
| CoreTrace, a visual 16-bit CPU simulator |
HN |
5 |
2026-08-13 |
| Run open-weight OCR, VLM and vision models behind one API |
HN |
5 |
2026-09-04 |
| Clawbernetes |
HN |
5 |
2026-02-20 |
| Symbolic regression as an MCP tool (SINDy and PySR, free, no install) |
HN |
5 |
2026-04-02 |
| TabPFN Scaling Mode |
HN |
5 |
2025-12-03 |
| Aurion OS, A 1.8MB OS with a browser, try it live (C/x86 ASM) |
HN |
5 |
2026-04-03 |
| Sub-microsecond (890 ns) trading execution research system |
HN |
5 |
2025-12-15 |
| Axiom |
HN |
5 |
2026-02-02 |
| Scope-structured arena memory for C, O(1) cleanup, no GC/borrow checker |
HN |
5 |
2026-04-15 |
| go-binsync |
HN |
5 |
2026-08-27 |
| TurboBench, the Compression Lie Detector, 100 Codecs, Daily Update |
HN |
5 |
2026-09-16 |
| Fixing LLM memory degradation in long coding sessions |
HN |
5 |
2025-11-27 |
| A new language for COBOL workloads, built on Go |
HN |
5 |
2025-11-05 |
| Building a full agentic harness around a 4B model is hard |
HN |
5 |
2026-08-19 |
| ClawMem |
HN |
5 |
2026-03-22 |
| On the edge of Apple Silicon memory speeds |
HN |
5 |
2026-01-17 |
| Anchor Engine |
HN |
5 |
2026-03-06 |
| m6502, a 6502 CPU for FPGAs and Tiny Tapeout |
HN |
5 |
2026-02-18 |
| I built a local Elixir/Python pipeline to curate 14,000 RAW photos |
HN |
5 |
2026-04-20 |
| Sipp |
HN |
5 |
2026-06-24 |
| Sofka |
HN |
5 |
2026-08-19 |
| I run 30B 22tok/s, 109tok/s not novel,6GB/16GB RAM overcoming llama.cpp |
HN |
5 |
2026-07-29 |
| Fast NF4 dequantization Triton kernel (1.41x faster than bitsandbytes) |
HN |
5 |
2026-07-15 |
| MiniVim a Minimal Neovim Configuration |
HN |
5 |
2026-02-24 |
| RamScout |
HN |
5 |
2025-12-08 |