LLM planner maps local hardware builds against open-weights performance benchmarks

A review of LLM planner, an open-source hardware and model matching tool aggregating 60 builds, 50 models, and power metrics to optimize local inference setups.

For founders and engineers building local AI pipelines, LLM planner is an indispensable directory that eliminates the tedious manual math of matching model parameter sizes to GPU VRAM limits. It is best for teams sizing on-premise workstations or edge deployment rigs who need to estimate throughput and power draw before purchasing hardware. Skip it if you rely entirely on hosted APIs or require guaranteed, locally verified latency figures. The tool is a highly organized aggregator of public benchmarks, not an active testing lab.

Methodology

This review evaluates LLM planner, observed in May 2026, based on the web application at llmrequirements.com and its public data repository on GitHub under Trenin-Labs/LlmRequirements. Because we are evaluating a metadata aggregator, our methodology focuses on verifying the integrity, sourcing, and structural utility of its database rather than running physical hardware tests. We analyzed the tool's mapping of 60 hardware configurations against 50 models, checking its cited sources, which include llama.cpp benchmark threads, Tom's Hardware, and community benchmarks. This review does not independently benchmark the physical hardware configurations or verify the exact tokens-per-second rates under varying software runtimes. It assesses the tool's utility as a planning instrument for technical decision-makers.

Bidirectional hardware and model matching

LLM planner features a bidirectional search interface. Users can input a specific hardware rig, such as an RTX 3090, Apple M3 Max, or Asus DGX Spark, to see which open-weights models fit within its memory limits. Alternatively, users can select a target model and quantization level (Q2, Q4, Q5, or Q8) to receive a list of compatible hardware builds. The tool ranks compatible models across functional categories like chat, coding, agents, and reasoning.

Granular power and cost metrics

In addition to memory compatibility, the tool aggregates estimated idle and active power draw in watts for each hardware configuration. This is a critical metric for continuous local operations where electricity costs can quickly eclipse initial hardware capital expenditure. It also provides multi-region pricing estimates for both new and used components, allowing builders to calculate the total cost of ownership.

Sourced performance reference points

Instead of presenting isolated performance numbers, the platform links its throughput estimates, covering decode and prompt-processing tokens per second, to 130 cited sources. These sources include public leaderboards, model cards, and community benchmark threads. Additionally, the tool indexes over 150 hardware review videos, providing direct links to external video evidence for each build configuration.

What is interesting

LLM planner succeeds by consolidating fragmented community knowledge. Before this tool, sizing a local LLM rig required cross-referencing GitHub issues, Reddit threads, and VRAM calculators to estimate if a model like Llama 3 70B would run acceptably on a dual-RTX 3090 setup. By compiling these variables into a single interface, the creator (totosse17) has built a functional calculator that saves hours of pre-purchase research.

However, users must treat the performance metrics as directional estimates rather than absolute guarantees. The tokens-per-second figures are compiled from disparate sources running different software backends, driver versions, and optimization libraries. A benchmark run on llama.cpp with specific CPU optimizations might not match the performance of an ExLlamaV2 implementation on the same GPU. The tool openly acknowledges this limitation, noting that it does not guarantee absolute best throughput.

Furthermore, the pricing data is highly volatile. While the tool provides multi-region used and new pricing, these figures are aggregated estimates rather than live inventory feeds. For builders in regions with highly volatile hardware markets, these numbers will serve only as a baseline. The inclusion of closed-source frontier models, like GPT-5.5 and Sonnet 4.6, as ceiling references is a helpful design touch, giving builders a clear performance anchor to decide if hosting locally is economically and operationally viable compared to calling an API.

Pricing

LLM planner is entirely free and open-source. The underlying dataset and codebase are hosted publicly on GitHub under the Trenin-Labs/LlmRequirements repository. Users can submit corrections or new benchmarks via GitHub issues. Pricing snapshot date: May 2026.

Verdict

LLM planner is a highly practical, community-driven directory that succeeds because it solves a highly specific, painful math problem for local AI builders. If you are tasking an engineering team with deploying local open-weights models for privacy, latency, or cost reasons, this tool should be your first stop to draft hardware budgets. It does not replace real-world testing on your specific software stack, but it dramatically narrows down the hardware configurations you should consider. We recommend using it to filter your options down to two or three builds before purchasing hardware for internal benchmarking.

What we would test next

In a future evaluation, we would like to test the variance between the tool's aggregated tokens-per-second claims and actual performance on a standardized local testbed. Specifically, we want to measure how closely the estimated active power draw matches real-world consumption during sustained batch inference workloads, as transient power spikes on multi-GPU setups often exceed nominal active watt ratings.

The investor read

The existence of LLM planner highlights a growing enterprise and developer push toward local, open-weights inference. As frontier-class open models close the capability gap with closed APIs, the bottleneck shifts from model quality to hardware economics. This tool signals that the market for local AI hardware optimization is maturing, but remains highly fragmented. For investors, the opportunity lies not in the directory itself, which is a bootstrapped utility, but in the underlying trend: enterprise teams are actively calculating the return on investment of on-premise hardware versus cloud APIs. Startups building software layers that automate this hardware-software co-design, or orchestrate local-to-cloud hybrid inference, are capturing a highly lucrative, capital-efficient segment of the AI infrastructure stack.

Pull quote: “By compiling these variables into a single interface, the creator (totosse17) has built a functional calculator that saves hours of pre-purchase research.”