whichllm finds best local LLMs for your hardware
The more useful story here is not the tool itself but its reception: whichllm, a CLI that auto-detects a user's GPU/CPU/RAM and ranks HuggingFace models by merged benchmark score, has grown from about 500 GitHub stars at its March 2026 launch to 4.8k stars and 264 forks by June, and an independent comparison by RunAIHome ranks it first among five local-LLM benchmarking tools. But a highly-voted Show HN thread (274 points) surfaces a pointed, credible critique: a top comment argues the tool actually recommends outdated models and misses newer ones, that installing and running a hardware-scanning, model-downloading CLI from an unfamiliar author carries real security risk compared to a browser-based alternative, and that both the project (down to its README) and the account's other activity look AI-generated rather than hand-written. None of this has been independently confirmed or refuted by the maintainer in what's publicly visible. For practitioners, the actionable lesson isn't about this one tool - it's a reminder to treat GitHub star counts and star-history charts as a popularity signal, not a trust signal, and to read a tool's actual source before running anything that touches your hardware and downloads multi-gigabyte files on its own.
The real value of this story for practitioners is a trust-verification lesson, not a tool recommendation: rapid GitHub star growth and favorable comparison-blog coverage do not, on their own, establish that a hardware-scanning CLI is safe or accurate - and this specific case has a credible, on-the-record community critique alongside its growth numbers that a quick look at the repo would not surface.
What happened
whichllm is an open-source Python CLI (Andyyyy64/whichllm on GitHub) that detects a user's GPU, CPU, and RAM, then ranks HuggingFace models using a merged score built from LiveBench, Artificial Analysis, Aider, Chatbot Arena ELO, and the Open LLM Leaderboard, with a stated recency-aware adjustment so older models cannot outrank newer ones on stale scores. Example output in the README shows Qwen/Qwen3.6-27B ranked first for an RTX 4090 with a score of 92.8 at roughly 27 tokens/second. The project launched in March 2026 and had grown to 4.8k GitHub stars and 264 forks by mid-June, with its latest release (v0.5.10) shipping June 11.
Independent review and community reception
RunAIHome's comparative review of five local-LLM benchmarking tools ranks whichllm first for the "what model should I run" question, pairing it with tools like LocalScore and llama-bench for hardware-specific throughput validation - but flags as a limitation that whichllm's scores come from merged community leaderboards rather than tests run on the user's own machine. Separately, a Show HN submission for the project drew 274 points and 63 comments; the top-voted reply argues three things: that the tool's actual recommendations skew toward outdated models and miss newer releases, that running a hardware-detecting, model-downloading CLI from an unfamiliar author on one's own machine carries real security risk relative to a browser-based hardware-spec calculator, and that the project's code, its README, and the submitting account's other comment history all read as AI-generated rather than human-authored. As of this writing, no maintainer rebuttal to that specific critique is visible in public reporting.
For practitioners
Treat this as a case study in evaluating developer tooling, not as an endorsement or a takedown: verify claims like a benchmark methodology or model-currency by reading the actual scoring logic (whichllm's README does publish its scoring formula and evidence-confidence tiers, which is a genuine transparency plus), and weigh GitHub star velocity against independent, critical community discussion rather than either alone. For any tool that inspects your hardware and pulls multi-gigabyte model files automatically, reviewing the source or at least the dependency list before running it is a reasonable baseline, regardless of star count.
What to watch
Whether the maintainer responds publicly to the outdated-model-ranking and AI-generated-content critiques; whether independent benchmarking (comparing whichllm's picks against fresh LiveBench/Arena data over time) supports or undercuts the recency claims in its marketing; and whether GitHub star growth continues at the same pace once the Show HN critique becomes more visible in search results.
Key Points
- 1whichllm grew from ~500 to 4.8k GitHub stars in three months and ranks first in an independent 5-tool comparison, per RunAIHome.
- 2A 274-point Show HN thread's top comment argues the tool recommends outdated models and that the project looks AI-generated ('vibecoded').
- 3The same critique flags real security risk in running an unfamiliar hardware-scanning, model-downloading CLI - star counts alone are not a trust signal.
Scoring Rationale
Backfilled with the tool's actual GitHub growth trajectory (500 to 4.8k stars), an independent comparative review (RunAIHome), and - most importantly - a credible, highly-voted Hacker News critique questioning both the tool's accuracy and its authenticity, none of which was in the original single-source writeup. This is a more complete and more useful practitioner story (evaluate before you trust and run) than the original uncritical feature recap. Held near the original level (6.3 to 6.0) reflecting that the added critique tempers rather than raises the story's substance - it remains a niche developer-tool item, not a technical or research milestone.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
