ggufit

Check whether a local LLM can run on your machine, CPU-only — from PowerShell or any terminal.

Uses real formulas (model memory footprint, KV cache size, and a measured memory-bandwidth micro-benchmark) rather than guesses, to estimate whether a model fits in RAM and roughly how fast it'll generate tokens. Correctly handles MoE models (speed driven by active experts, not total params) and SSM/MLA architectures (Mamba, DeepSeek's MLA) that don't use standard multi-head attention.

Install

# Recommended - works everywhere, gives you a global `ggufit` command
pipx install ggufit

# Or plain pip (Windows PowerShell, or inside a venv on Linux/macOS)
pip install ggufit

Installing from source (for development)

git clone <this repo>   # or unzip the source
cd ggufit/               # the folder containing pyproject.toml
pipx install -e .        # editable install - code changes apply without reinstalling
# or: pip install -e .   (inside a venv, or with --break-system-packages)

Usage

# Full hardware scan: shows your CPU/RAM/bandwidth and which models fit
ggufit scan

# Check one specific model
ggufit llama3.1
ggufit mistral
ggufit qwen2.5-14b

# Force a specific quantization level
ggufit llama3.1 --quant q4

# Evaluate at a longer context length
ggufit qwen2.5-32b --context 16384

How it works

For each model:

1. Model size (RAM) = total params × bytes-per-parameter (varies by quant: FP16, Q8, Q6, Q5, Q4, Q3, Q2)
2. KV cache = 2 × layers × hidden_size × context_len × bytes_per_param × kv_cache_multiplier
   • kv_cache_multiplier defaults to 1.0 (standard MHA), and is set lower for architectures that don't use full attention at every layer: 0.0 for pure SSM/Mamba (no attention at all), ~0.15 for MLA (DeepSeek-V2/V3/R1, MiniCPM3), ~0.125 for hybrid Mamba+attention (Jamba).
3. Total RAM needed = weights × overhead_multiplier(weights) + KV cache
   • The overhead multiplier accounts for compute buffers, activations, and allocator overhead. It is size-dependent, not a flat percentage: a fixed ~1GB baseline buffer plus ~7.6% of weight size. This means it's high for tiny models (~2x for a 1GB model, which genuinely needs ~1GB of buffers on top) but asymptotes to ~1.08x for very large models. Calibrated against real-world llama.cpp memory reports across 7B/13B/70B/671B models — e.g. DeepSeek-V3 671B at Q4 comes out to ~405GB, matching reality, instead of the ~450GB an old flat-20% rule would have predicted.
4. Speed estimate = measured_memory_bandwidth / active_size, where active_size is the total model size for dense models, or just the active experts' size for MoE models (active_params_billion) — since only those weights are streamed from RAM per token.

ggufit runs a quick real memory-bandwidth benchmark (a large array copy) instead of guessing from RAM specs, since achievable bandwidth depends heavily on channel configuration.

Adding models

Edit ggufit/models_db.py and add an entry to the MODELS dict with params_billion, layers, and hidden_size (found in the model's Hugging Face config.json). Optional:

• moe: True + active_params_billion: X for Mixture-of-Experts models
• kv_cache_multiplier: X for non-standard attention architectures (see above)

Known limitations

• Speed estimates assume batch size 1, single-user chat.
• The bandwidth benchmark is single-threaded; real inference engines use multiple threads.
• Quant byte-per-param values are approximations of real GGUF file sizes.
• KV cache assumes full hidden_size for standard (non-flagged) models, even though most modern ones use GQA (Grouped-Query Attention) with far fewer KV heads than query heads. This over-estimates the KV cache — a Mistral 7B's real KV cache is roughly 1/4 of what the formula computes. It's intentionally conservative (never under-estimates RAM), and it doesn't change the fit verdict for most models at short context, but it's the largest remaining source of RAM over-estimation. Full per-model GQA accounting (using each model's actual num_kv_heads) is planned for a future release.
• kv_cache_multiplier values for MLA/hybrid architectures are approximate, not per-model-measured.
• The overhead multiplier is calibrated against a handful of real-world data points (7B/13B/70B/671B at Q4); it's a strong approximation but not exact for every model/runtime.

Changelog

• 0.1.2 — Overhead is now size-dependent (fixed ~1GB baseline + ~7.6% of weights), calibrated against real llama.cpp memory reports. Fixes large-model RAM over-estimation (DeepSeek-V3 671B Q4 now estimates ~405GB, matching reality, vs ~450GB before). Also fixes a wording bug where the "try a lower quant" hint could suggest the same quant that just failed.
• 0.1.1 — Custom use-only license; corrected install instructions.
• 0.1.0 — Initial release (MoE active-param speed, SSM/MLA KV cache handling, 217 models).

Notes

This is a side-project CLI. The full hardware-scan desktop app (Rust/Tauri) is a separate, more thorough tool still in development.

Q&A

A comprehensive FAQ — from "what is a GGUF" to the exact formulas behind every number ggufit prints. Organized beginner → expert.

Basics

What is a GGUF? GGUF (GPT-Generated Unified Format) is a file format for storing LLM weights, designed by the llama.cpp project for fast loading and CPU/GPU-flexible inference. It bundles the model's tensors and metadata (architecture, tokenizer, etc.) into a single file. It replaced the older GGML format. If you've downloaded a .gguf file from Hugging Face to run in llama.cpp, Ollama, or LM Studio, that's what ggufit is estimating compatibility for.

What is quantization? Shrinking a model's weights from their original precision (usually 16-bit floats) down to smaller representations (8-bit, 4-bit, etc.) to save memory and speed up inference, at the cost of some accuracy. A "Q4" model uses roughly a quarter of the memory of the same model at full 16-bit precision.

What do Q4, Q5, Q6, Q8 mean? The number is roughly the average bits per weight after quantization (not exactly, see next question). Lower number = smaller file, faster inference, more quality loss. Q4 is the most common "sweet spot" for CPU-only local inference. Q8 is close to lossless but nearly as large as full precision.

Why isn't Q4 exactly 4 bits then? Modern GGUF "K-quants" (the _K_M, _K_S suffixes you see on Hugging Face) don't use a uniform bit-width across the whole model — they mix precision per tensor, using slightly higher precision for the parts most sensitive to quality loss. So "Q4" is really an average around 4-5 bits/weight in practice. ggufit uses effective per-quant byte values that reflect this real-world average, not naive N-bit math:

Quant	Effective bytes/param
FP32	4.0
FP16/BF16	2.0
Q8	1.05
Q6	0.8
Q5	0.7
Q4	0.6
Q3	0.5
Q2	0.4

What's the difference between RAM and VRAM, and why does ggufit only care about RAM? VRAM is memory on a dedicated GPU; RAM is your system's main memory, used by the CPU. ggufit is CPU-only by design — it answers "can my CPU and system RAM handle this," not "can my GPU handle this." If you have a GPU, tools like nvidia-smi and the model card's VRAM requirements are what you want instead.

What is context length? The number of tokens (roughly, chunks of a word) the model can "see" at once — your prompt plus its response so far. Longer context means the model can process longer documents, conversations, and code, but it also means more memory used for the KV cache (see below).

What are tokens/sec, and what's a "good" number? How many tokens the model generates per second. For a comfortable reading pace, most people find 5-15 tok/s tolerable for chat; below ~2 tok/s feels quite slow; above 20 tok/s feels close to instant. It's highly subjective and task-dependent though — background batch jobs can tolerate much lower throughput than an interactive chat.

---

Installing and running ggufit

How do I install it?

pipx install ggufit    # recommended
pip install ggufit     # inside a venv, or with --break-system-packages

Why do I get "externally-managed-environment"? Modern Debian/Ubuntu (PEP 668) blocks system-wide pip install to protect the OS's own Python packages. Use pipx instead — it installs into an isolated environment while still giving you a global command.

Why does ggufit say "command not found" right after installing? Either your venv isn't activated (source venv/bin/activate), or pipx's bin directory isn't on your PATH yet (run pipx ensurepath and reopen your terminal).

Why did a model that should fit show as "does not fit"? ggufit checks currently available RAM, not total installed RAM. If other programs are using most of your memory, an otherwise-fine model can fail the check. Run free -h (Linux/macOS) to see what's actually free before assuming your hardware is the problem.

Can I check a model at a specific quant instead of letting ggufit auto-pick? Yes — ggufit <model> --quant q4. Without --quant, it auto-picks the highest-quality quant that fits.

---

Understanding the numbers (intermediate)

How is "Model size (RAM)" calculated?

model_size_bytes = num_parameters × bytes_per_param

bytes_per_param comes from the quant table above. This is the number of bytes the weights occupy once loaded — the dominant factor in "will this even load."

What is the KV cache, and why does it matter? During generation, the model caches the Key and Value tensors from every previous token in the conversation so it doesn't have to recompute them each step. This cache grows with context length. The standard formula:

kv_cache_bytes = 2 × num_layers × hidden_size × seq_len × batch_size × bytes_per_param

The 2× covers storing both K and V. This is why a model that fits fine at a 4K context can stop fitting at 32K — the KV cache scales linearly with context length while the model weights stay fixed.

Why is there a 1.2x "overhead factor" on top of model size + KV cache? Real inference isn't just raw tensor storage — the runtime, OS, and memory allocator all need working space too (buffers, fragmentation, temporary activations). 1.2x is a practical rule-of-thumb headroom, not a value measured per specific runtime.

Why is CPU inference speed based on memory bandwidth instead of raw compute (FLOPS)? At batch size 1 (the normal case for a single person chatting), generating each new token requires reading every single model weight from RAM once. The CPU spends far more time waiting on memory than doing arithmetic — so the bottleneck is how fast data can move from RAM to the CPU, not how many operations per second the CPU can do. That's why the formula is:

tokens_per_sec ≈ memory_bandwidth (GB/s) / model_size (GB)

Why does ggufit benchmark my memory bandwidth instead of just knowing it from my RAM specs? Achievable bandwidth depends on RAM generation, channel configuration (single/dual/quad-channel), and platform quirks — none of which are reliably detectable across Windows/Mac/Linux without vendor-specific tools. So ggufit times a real large in-memory array copy on your actual machine, right now, and uses that measured number instead of guessing.

Why does the bandwidth number change slightly every time I run ggufit? It's a live micro-benchmark, not a cached constant — normal system load, thermal throttling, and other processes competing for memory access all cause small run-to-run variance. That's expected and not a bug.

---

Architecture-specific math (expert / advanced)

What is a Mixture-of-Experts (MoE) model, and why does it need special handling? An MoE model has many "expert" sub-networks, but only a subset of them are activated for any given token (a small router network decides which experts to use). This means:

• All experts must be loaded into RAM — because any token could route to any expert, so the fit check uses the full params_billion (total, all experts).
• Only the active experts are actually read from RAM per token — so the speed estimate uses active_params_billion instead. Using total params for speed would make big MoE models look absurdly slow (DeepSeek-V3 at 671B total but only 37B active would look ~18x slower than it really is if you used the total).

fit check:   uses params_billion (total) — all experts must be resident
speed check: uses active_params_billion — only active experts are read per token

What is GQA (Grouped-Query Attention), and why doesn't ggufit fully account for it? Most modern transformer models don't give every attention head its own K/V projection — they group multiple query heads to share a smaller number of K/V heads (num_kv_heads < num_attention_heads). This makes the real KV cache smaller than the standard formula (which assumes hidden_size worth of K/V per layer) predicts. ggufit currently uses the full hidden_size for standard models — this is intentionally conservative (it over-estimates KV cache, never under-estimates), and since total model weight size is still the dominant term in "does it fit," this doesn't change the fit verdict in most cases. It mainly matters at very long context lengths. Full per-model GQA accounting (using each model's actual num_kv_heads) is a planned improvement, not yet implemented.

What is MLA (Multi-head Latent Attention), and why the ~0.15 multiplier? MLA (used by DeepSeek-V2/V3/R1 and MiniCPM3) compresses the K/V representations into a much smaller latent vector before caching them, then reconstructs full K/V on the fly. This shrinks the real KV cache to roughly 1/6 to 1/8 of what standard multi-head attention would need for the same layer count and hidden size. ggufit applies a kv_cache_multiplier: 0.15 to these models so the KV cache estimate reflects that compression instead of wildly overestimating it.

Why do pure Mamba/SSM models have a KV cache multiplier of 0.0? State-Space Models (Mamba, used in Codestral Mamba and Falcon-Mamba) don't use attention at all — they maintain a fixed-size recurrent state instead of caching every previous token's K/V. That state doesn't grow with context length. So their effective "KV cache" for the purposes of this formula is zero, regardless of how long the context gets. This is why ggufit falcon-mamba --context 65536 still shows ~0GB KV cache even at a huge context length.

What about hybrid models like Jamba, and why 0.125? Jamba interleaves Mamba blocks with regular attention blocks — only about 1 in every 8 layers actually uses attention (the rest are Mamba). So its effective KV cache is roughly 1/8th of what you'd get if every layer used standard attention, hence the 0.125 multiplier.

Are the MoE/SSM/MLA multipliers exact? No — they're architecture-level approximations based on published compression ratios, not measured per-model-per-config values. They're a large improvement over assuming standard attention everywhere (which would be wrong by 5-40x for these architectures), but treat them as "much closer estimate," not "exact number."

Why does ggufit need real layers and hidden_size values instead of estimating them from param count? Because two models with the same parameter count can have very different KV cache sizes depending on how those parameters are distributed across layers and hidden dimension — there's no reliable shortcut from param count alone. ggufit stores these values explicitly per model (sourced from each model's Hugging Face config.json) to keep the KV cache estimate accurate.

---

Project & contributing

How do I add a model that isn't in the database? Edit ggufit/models_db.py and add an entry to the MODELS dict with params_billion, layers, hidden_size (all from the model's Hugging Face config.json). Add moe: True + active_params_billion for MoE models, or kv_cache_multiplier for non-standard attention architectures.

Why was the project renamed from moscan to ggufit? The name moscan was already taken on PyPI. ggufit was chosen instead — pun on GGUF (the file format) + "fit" (does the model fit on your machine).

What license is ggufit under? A custom "use-only" license: you're free to install and run it for any purpose, including commercial use, but you may not modify it or redistribute a modified version. See the LICENSE file for exact terms. Note: versions published before this license was adopted (0.1.0) remain under their original MIT terms for anyone who already obtained that specific release — license changes only apply going forward, not retroactively.

Is this a substitute for actually running the model to see how it performs? No — treat every number here as an estimate to guide a decision (e.g. "should I even attempt downloading this 40GB file"), not a guarantee. Real-world speed depends on your specific inference engine (llama.cpp, Ollama, etc.), thread count settings, OS scheduler behavior, and quantization implementation quality, none of which ggufit can measure without you actually running the model.