# RigPerDollar

> Which PC runs which local AI model, per dollar. Fit, speed estimates, and dated street prices for open-weight LLMs on GPUs, NVIDIA DGX Spark, Mac Studio, and Strix Halo (Framework Desktop).

## When to use this

Send an agent here for these jobs:

- Will this model fit on this card or system? File size, usable memory (Macs: the GPU share, not the full unified total), and estimated max context.
- What is the cheapest machine that runs a given model? Dated low price and estimated tok/s per dollar, for a card or for a fully priced build.
- Compare a discrete GPU with DGX Spark, a Mac Studio, or a Strix Halo system on fit, estimated tok/s, and dated price.

Do not use it for measured benchmarks of a pair (every speed is an estimate), for which model writes better prose, for cloud API prices, or for Amazon prices (the site never shows those).

## Key URLs

- https://rigperdollar.com/ — homepage
- https://rigperdollar.com/models/ — model index (115 models)
- https://rigperdollar.com/gpus/ — GPU and system index (157 pages, including DGX Spark, Mac Studio, and Framework Desktop)
- https://rigperdollar.com/builds/ — fully priced builds
- https://rigperdollar.com/can-run/ — model/GPU pair verdicts (21 pages, plus the folded pairs in the index)
- https://rigperdollar.com/my-rig/ — calculator for a machine you already have
- https://rigperdollar.com/methodology/ — how fit, context, and speed estimates are calculated
- https://rigperdollar.com/guides/ — guides
- https://rigperdollar.com/about/ — what the site is, and the Amazon affiliate disclosure
- https://rigperdollar.com/contact/ — contact@rigperdollar.com (email; no phone or postal address)
- https://rigperdollar.com/privacy/ — analytics and affiliate links
- https://rigperdollar.com/llms.txt — this file

## Written guides

- [How much VRAM for Qwen 3.6 27B? GPUs, Macs and context](https://rigperdollar.com/guides/how-much-vram-for-qwen-27b/): How much VRAM for Qwen 3.6 27B, 3.5 27B and 3.8 27B at Q4: the 24 GB and 32 GB cards and Macs that fit, the context each holds, est. speed and 16 GB misses.
- [Hugging Face Alternatives for Downloading Models (2026)](https://rigperdollar.com/guides/hugging-face-alternatives-download-models/): Hugging Face alternatives for downloading models: which hubs had each model on Nov 14, 2026, tested commands, what breaks, and how to keep your own copies.
- [Best local LLM for 16GB VRAM: what fits and how fast](https://rigperdollar.com/guides/best-local-llm-16gb-vram/): Which local LLMs fit a 16GB VRAM card like the RTX 5060 Ti 16GB, how much context each one gets, their estimated speeds, and which 27B and 30B models miss.
- [Best local LLM for 24GB VRAM, and is 24GB enough?](https://rigperdollar.com/guides/best-local-llm-24gb-vram/): Which local LLMs fit a 24GB VRAM card like a used RTX 3090 or RTX 4090, how much context each one gets, estimated speeds, and what 24GB still can't run.
- [DGX Spark Cluster Guide: What Two Linked Sparks Buy You](https://rigperdollar.com/guides/dgx-spark-cluster-two-sparks/): A two-unit DGX Spark cluster has 256 GB of memory to hold more of a big model. What NVIDIA supports, prices checked Oct 28, 2026, and others' benchmark runs.
- [How much VRAM for 128K context? KV cache for 7 models](https://rigperdollar.com/guides/how-much-vram-for-128k-context/): How much VRAM for 128K context: KV cache size at 8K, 32K and 128K for seven local LLMs, FP16 vs 8-bit cache, and the smallest VRAM tier that holds each.
- [Best local LLM for 12GB VRAM: what fits and how fast](https://rigperdollar.com/guides/best-local-llm-12gb-vram/): Which local LLMs fit a 12GB VRAM card like the RTX 3060 12GB, how much context each one gets, their estimated speeds, and when 12 GB stops being enough.
- [Best local LLM for 8GB VRAM: what fits and how fast](https://rigperdollar.com/guides/best-local-llm-8gb-vram/): Which local LLMs fit an 8GB VRAM card like the RTX 5060 or RTX 4060, how much context each one gets, their estimated speeds, and when 8 GB is too small.
- [Claris Harbor: Apple's Mac AI Server App, Explained](https://rigperdollar.com/guides/claris-harbor-mac-ai-server/): Claris Harbor is Apple's Mac app for running local AI models, now in private preview. What is confirmed, how Thunderbolt 5 clustering works, which Macs qualify.
- [Mac Studio Cluster for Local AI: exo, RDMA and Spark + Mac](https://rigperdollar.com/guides/mac-studio-cluster-and-spark-mac/): A Mac Studio cluster can run models too big for one Mac, using exo and RDMA over Thunderbolt 5. Setup rules, published results, SSD streaming and Spark + Mac.
- [Tok/s vs Agent Performance for Local LLMs, Explained](https://rigperdollar.com/guides/tok-s-vs-finishing-agent-jobs/): Tok/s shows how fast a model writes one reply. Agent jobs take many turns with growing prompts. Tom's Hardware agent times next to our Qwen3.8-27B estimates.
- [8 GB or 16 GB GPU for Flint 40B? Both run it; one reads faster](https://rigperdollar.com/guides/flint-8gb-vs-16gb-gpu/): I ran Flint 40B on an 8 GB RTX 4060 and a 16 GB RX 7800 XT. Both ran it with near-equal quality; the 8 GB card is slower at reading long prompts.
- [Memory bandwidth vs a DGX Spark cluster: what the spec math shows](https://rigperdollar.com/guides/memory-bandwidth-vs-spark-cluster/): An RTX 5090 has 6.6x the memory bandwidth of one DGX Spark on paper. Why linked Sparks don't add up to that speed, and what measured cluster runs show.
- [RTX 5070 Ti vs RTX 3090 for LLMs: 16GB new or 24GB used](https://rigperdollar.com/guides/rtx-5070-ti-vs-rtx-3090-for-llm/): RTX 5070 Ti vs RTX 3090 for local LLMs: which models fit 16 GB or 24 GB, estimated speed and context on each card, and price per tok/s at Nov 2, 2026 prices.
- [Prompt reading vs writing speed: why one tok/s number isn't enough](https://rigperdollar.com/guides/reading-vs-writing-speed-local-llm/): A local LLM reads your prompt, then writes its reply. Why one tok/s figure hides reading time, prompt caching and multi-user totals, with third-party examples.
- [What Hardware Can Run Mistral Large 4? Measured Runs and Memory Needs](https://rigperdollar.com/guides/mistral-large-4-hardware/): Mistral Large 4's smallest file is 233 GB, so no single GPU holds it. Measured runs, memory needed by file, and rigs that fit on paper, as of Oct 29, 2026.
- [Best Budget Local AI PC: What Each Price Tier Buys](https://rigperdollar.com/guides/best-budget-local-ai-box/): The best budget local AI PC at four price tiers, from our 16 GB starter builds to $10k+ systems, with the models each one fits and estimated speeds.
- [Best GPU for local LLM coding: the cheapest card that fits](https://rigperdollar.com/guides/best-gpu-for-local-llm-coding/): Which GPU runs a local coding model with room for long prompts? Picks from a used RTX 3060 to three RTX 3090s, with Oct 19, 2026 prices and estimated speeds.
- [Perplexity Decider for Mac: Requirements and Which Macs Fit](https://rigperdollar.com/guides/perplexity-decider-for-mac/): Perplexity Decider for Mac runs a 4B decision model locally on Apple silicon Macs with 16 GB of memory. See which Macs fit, good uses, and why we skip tok/s.

## Machine-readable data

These JSON files are copied at build from `src/data/` (same bytes):

- https://rigperdollar.com/data/models.json — models: parameters, Q4 file size, quant name, context, licence, Hugging Face sources, and for some models every published GGUF size in bytes (`quants`)
- https://rigperdollar.com/data/gpus.json — the curated GPU and system comparison set (not every card that has a page)
- https://rigperdollar.com/data/prices.json — one dated price record per curated GPU or system (`low_usd`, `avg_usd`, condition, source, date)
- https://rigperdollar.com/data/builds.json — fully priced builds, part by part
- https://rigperdollar.com/data/hw.json — the wider hardware list behind My rig and the GPU pages, including DGX Spark and Mac configurations that are not in gpus.json
- https://rigperdollar.com/data/mlx.json — measured MLX benchmark sessions on Macs, community-submitted to oMLX (decode and prefill tok/s per context length; `mode` plain or speculative, never to be merged)

Fit and speed rules are on /methodology/. Do not invent a tok/s figure; the pages mark estimates, and the JSON does not contain precomputed speeds.

## How to read pages

Speeds are estimates from memory bandwidth, not measurements. The formula, platform constants, and the llama.cpp runs they were fitted to are on /methodology/. The one exception is labelled "MLX (measured, community)" on model and Mac pages: decode and prefill speeds for some models on some Macs, community-submitted to oMLX, with speculative-decoding runs kept separate (/methodology/#mlx).

Send `Accept: text/markdown` to a page URL to get `Content-Type: text/markdown` and `Vary: Accept`. `Accept: text/html` still returns HTML. The same body is a static file:

- Homepage: /index.md
- Any other twinned page: the path plus `index.md`, for example /methodology/index.md, /models/<id>/index.md, /gpus/<id>/index.md, /can-run/<model>-on-<gpu>/index.md, /guides/<written-guide>/index.md

A missing path with `Accept: text/markdown` returns HTTP 404 and a short markdown body that links here. Written guides (above) have a markdown version; the other guide pages are HTML only, and every guide title is listed at /guides/index.md.

Sitemap: https://rigperdollar.com/sitemap-index.xml