Local AI Report #3 — Best Small LLMs for 8 GB and 16 GB RAM Laptops

Issue #3 · August 25, 2026 · Last updated: August 25, 2026 · Jakub Rusinowski

An 8 GB laptop leaves about 6.4 GB for a model, and a 16 GB laptop about 12.8 GB. Here is what actually fits, how much context you get, and how fast it runs with no discrete GPU.

TL;DR

  • An 8 GB laptop gives a model about 6.4 GB, and a 16 GB laptop about 12.8 GB — the rest belongs to the OS and everything else you have open.
  • On 8 GB, a 7B–8B model at Q4 does not fit once you ask for real context. The honest ceiling is a 4B-class model at 16K.
  • On 16 GB, a 14B at Q4 fits at 16K context — 8.94 GB of weights plus cache — but it is the slowest thing on this page.
  • Speed is decided by memory bandwidth, not your CPU. A laptop runs at 38–120 GB/s against a discrete card's 300–1000+.
  • One memory stick instead of two halves your speed: DDR5-5600 is 44.8 GB/s single-channel and 89.6 GB/s dual. For many readers this is a bigger win than changing model.
  • Every speed figure here is estimated, published as a range, and the range is wide because we hold zero CPU benchmarks of our own.

What 8 GB and 16 GB actually leave you

Your laptop does not hand the whole of its memory to a model. The operating system, your browser and everything else already resident take a share, and committing the remainder to a model is what turns a working machine into a swapping one.

We size the budget as a fraction of installed memory rather than by subtracting an "OS footprint", because the footprint question does not have one answer. On Windows, the figure in Task Manager labelled In use and the figure a vendor quotes as "used" are different quantities: the larger one usually includes the standby cache, which holds cached file data and is reclaimable on demand. A model can take that memory back. It cannot take back what is genuinely committed.

Read *Available*, not *In use*. Available is the number that tells you what a model can have.

Installed RAMPlatformBudget for the modelBasis
8 GBWindows / Linux6.40 GB80% of installed
16 GBWindows / Linux12.80 GB80% of installed
8 GBmacOS6.00 GB75% of unified memory
16 GBmacOS12.00 GB75% of unified memory

The macOS figure is lower because the GPU's share of unified memory is capped by the system. That cap is adjustable, so 75% is a sensible default rather than a hard wall.

This is system RAM, not VRAM. If you have a discrete graphics card, its dedicated memory is the number that matters instead and this page does not apply — see Best LLMs for 8 GB VRAM or Best LLMs for 16 GB VRAM. A discrete 8 GB card runs models this page rules out, because its bandwidth is five to ten times a laptop's.

The number in the model name is not the number that matters

A model's parameter count tells you how much memory it occupies. It does not tell you how fast it runs. Those are two different quantities and conflating them is the most common mistake in this category.

Residency is every parameter, because all of them must be in memory. Per-token traffic is only the parameters actually read to produce one token. For an ordinary dense model these are the same. For a Mixture-of-Experts model they are not: only a fraction of the experts activate per token, so a large MoE can occupy a great deal of memory and still generate quickly.

That is why the tables below carry two different columns, and why a 12B and a 26B-A4B are not comparable on size alone. The full arithmetic is at our methodology page, and what quantisation actually costs you covers the other half of the trade.

What fits in 8 GB of RAM

Every row computed at Q4_K_M with an f16 KV cache, against the 6.40 GB budget. "Max context" is the largest standard context length whose total footprint still fits.

ModelWeightsMax context that fitsTotal at that contextKV basis
Qwen 3.5 0.8B0.48 GB64K5.58 GBestimated
Qwen 3.5 2B1.21 GB32K4.96 GBestimated
Llama 3.2 3B1.94 GB16K4.62 GBmeasured
Phi-4 Mini2.29 GB16K5.24 GBmeasured
Gemma 3 4B2.42 GB16K5.50 GBmeasured
Qwen 3.5 4B2.42 GB16K5.03 GBestimated
Gemma 4 E2B3.08 GB16K5.89 GBestimated
Gemma 4 E4B4.83 GB4K6.20 GBestimated

The practical pick is a 4B-class model at 16K context. Llama 3.2 3B and Phi-4 Mini leave the most headroom; Gemma 3 4B is the largest that still holds 16K comfortably.

Do not pull these on 8 GB

ModelWeightsWhy not
Qwen 3.5 9B5.43 GBFits on paper, leaves under 1 GB for cache — no usable context
Gemma 4 12B7.25 GBExceeds the budget before any context
Phi-4 14B8.45 GBExceeds the budget by 2 GB
Qwen 3 14B8.94 GBExceeds the budget by 2.5 GB
GPT-OSS 20B12.07 GBNearly double the budget

The widely-repeated claim that "8 GB runs a 7B" is where most articles go wrong. A 7B–8B at Q4 is roughly 4.6–5.4 GB of weights alone; add a KV cache for anything past a couple of thousand tokens and you are over. It loads, then it swaps.

What fits in 16 GB of RAM

Same method, against the 12.80 GB budget.

ModelWeightsMax context that fitsTotal at that contextKV basis
Qwen 3.5 0.8B0.48 GB128K9.87 GBestimated
Qwen 3.5 2B1.21 GB64K7.91 GBestimated
Llama 3.2 3B1.94 GB64K10.25 GBmeasured
Phi-4 Mini2.29 GB64K11.68 GBmeasured
Qwen 3.5 4B2.42 GB64K10.46 GBestimated
Gemma 3 4B2.42 GB64K12.34 GBmeasured
Gemma 4 E2B3.08 GB64K11.93 GBestimated
Gemma 4 E4B4.83 GB32K10.19 GBestimated
Qwen 3.5 9B5.43 GB32K11.07 GBestimated
Gemma 4 12B7.25 GB16K10.66 GBestimated
Qwen 3 14B8.94 GB16K12.42 GBmeasured
Phi-4 14B8.45 GB16K12.61 GBmeasured

16 GB is where a laptop becomes genuinely useful. A 9B at 32K handles summarising and drafting without feeling cramped. The 14B models fit, but read the speed section before choosing one — they are three to four times slower than a 4B on the same machine, and on a laptop that is the difference between usable and abandoned.

Still nothing here for GPT-OSS 20B (12.07 GB), Gemma 4 26B-A4B (15.70 GB) or Qwen 3.5 27B (16.30 GB). Those need more memory than a 16 GB machine can give them.

For workload-specific picks once you know your budget, see general assistant, summarisation, translation and RAG over your own files.

How fast will it actually be?

Generating a token means reading the model's active weights out of memory. Every token. So the ceiling is arithmetic:

> tokens/sec = memory bandwidth × efficiency ÷ bytes read per token

Your CPU's core count barely enters into it. This is why a faster processor with the same memory does not help much, and why a second memory stick does.

What your laptop's memory actually delivers

ConfigurationPeak bandwidthHow it is derived
DDR4-2400, dual channel38.4 GB/s2 × 8 B × 2400 MT/s
DDR4-3200, dual channel51.2 GB/s2 × 8 B × 3200 MT/s
DDR5-5600, single channel44.8 GB/s1 × 8 B × 5600 MT/s
DDR5-5600, dual channel89.6 GB/s2 × 8 B × 5600 MT/s
LPDDR5X-7500, soldered120 GB/s128-bit bus × 7500 MT/s
Apple M1 (unified)68 GB/s
Apple M4 (unified)120 GB/s

For comparison, a mid-range discrete card is 300–500 GB/s and a high-end one exceeds 1,000. That gap, not the CPU, is why laptop inference is slower.

Estimated generation speed

Every figure below is ESTIMATED. We publish a range, not a point, and the range is wide on purpose — see the methodology note at the foot of this page.

ModelGB read per tokenDDR4-3200 dual · est. tok/sDDR5-5600 dual · est. tok/sApple M4 · est. tok/s
Qwen 3.5 2B1.5811–2820–4827–65
Llama 3.2 3B2.417–1813–3217–42
Phi-4 Mini2.836–1511–2715–36
Gemma 3 4B2.996–1511–2614–34
Gemma 4 E2B3.585–129–2112–28
Gemma 4 E4B5.403–86–148–19
Qwen 3.5 9B6.043–75–137–17
Gemma 4 12B7.902–64–105–13
Qwen 3 14B9.612–53–84–11

Roughly: anything under about 5 tok/s is painful for long output and fine for short answers. Above 15 feels conversational.

We are not publishing time-to-first-token. Our model for it is wrong by around an order of magnitude on hardware that is not an NVIDIA card, and a wrong TTFT is worse than none.

Check your memory channels before you change model

Single-channel memory is the most common unforced error on a laptop, and the arithmetic above shows exactly what it costs: DDR5-5600 delivers 44.8 GB/s on one stick and 89.6 GB/s on two. Exactly double. Since generation is bandwidth-bound, so is your token rate.

Many machines ship with one stick in a two-slot chassis. Adding a second matched stick is usually the cheapest performance change available, and it is a bigger win than dropping from a 4B to a 2B.

Check what you have:

Note the trade: soldered LPDDR5X in a thin-and-light is fast (120 GB/s) but cannot be upgraded. A socketed machine is slower out of the box and fixable.

Platform setup details are on Windows, macOS and Linux.

Context: the 256K number is not for you

Model cards advertise context windows of 128K or 256K tokens. On a laptop those numbers are unreachable, because the KV cache — the model's memory of the conversation so far — grows linearly with context and comes out of the same budget as the weights.

The cache cost per 1,024 tokens for the models above runs from about 0.067 GB (Qwen 3.5 0.8B) to 0.210 GB (Phi-4 14B). At 64K tokens that is between 4 GB and 13 GB — on its own, before the weights.

That is what the "max context that fits" column is doing: it is the largest standard length where weights plus cache plus runtime overhead still clear the budget.

The one-line fix: quantise the cache. Setting OLLAMA_KV_CACHE_TYPE=q8_0 roughly halves cache memory and buys back context. It costs some quality; we have not measured how much, so we are not going to put a number on it.

One caveat we would rather state than hide: the "KV basis" column marks whether the cache figure is computed from a model's published attention configuration (measured) or inferred from its parameter count (estimated). The inference is not reliable — across models where we can check it against the real configuration, its average error is about 51%. Treat every estimated row as approximate, and prefer a measured row when choosing between two similar models.

The limits — what this will not do for you

A local model on a laptop is genuinely useful for a narrow band of work. It is not a substitute for a frontier model, and the gap is largest exactly where people most want it to be small.

1. Long documents are slow at the start, not just during output. Reading a 4,000-word document is several thousand tokens of processing before the first word appears. On a laptop that wait is measured in minutes, not seconds. 2. A 4B model is not a 400B model. Expect competent summarising, rewriting and short answers. Do not expect reliable multi-step reasoning, current facts, or code you can ship unread. 3. Quality drops with quantisation and we are not quantifying it here. Q4 is the standard trade-off; below Q4 the loss becomes obvious. See quantisation explained. 4. Battery and heat are real. Sustained generation runs the machine hard. Fanless laptops throttle within minutes. 5. We hold zero CPU benchmarks. Every speed figure on this page is a model output, not a measurement — see the note below. 6. Context is far shorter than advertised, per the section above. 7. We are not publishing download sizes. The figure a model registry reports is the only authority, and we could not read it from our build environment. We would rather leave a gap than print a number we did not check. 8. An NPU will probably not help. Current local runtimes largely target CPU and GPU; a Copilot+ NPU is excellent at small always-on models and a poor general LLM engine. 9. A cheap discrete GPU beats all of this. Even an entry-level card with 8 GB of VRAM outruns any CPU configuration here, because its bandwidth is several times a laptop's — see our 4 GB VRAM guide and what 32 GB of VRAM buys if you are weighing an upgrade against running in the cloud.

Apple Silicon: same rules, better hardware

Apple Silicon runs the same arithmetic with two structural advantages. Memory is unified, so there is no separate VRAM pool to fill and no copying between them. And bandwidth is high for the class: 68 GB/s on an M1 and 120 GB/s on an M4, against 38–90 GB/s for typical socketed laptop memory.

The offsetting constraint is the system's cap on how much unified memory the GPU may address, which is why we budget 75% rather than 80%. On a 16 GB Mac that is 12.00 GB.

An 8 GB Apple laptop is still an 8 GB machine. It gets 6.00 GB and the same 4B-class ceiling as any other 8 GB laptop — the bandwidth advantage makes it faster, not roomier.

Setup specifics are on Local AI on macOS.

Methodology, and how much to trust these numbers

Memory figures are exact arithmetic and we stand behind them. Weights are the parameter count times the quantisation's bits-per-weight. The KV cache is computed from each model's attention configuration where we have it, and inferred from parameter count where we do not — the "KV basis" column says which. Total footprint adds a fixed runtime overhead. The same functions produce these numbers as produce every other memory figure on this site, so a page here cannot disagree with the VRAM calculator or with our methodology page.

Speed figures are estimates, and here is precisely why they are ranges. Throughput is bandwidth × efficiency ÷ bytes-per-token. Bandwidth and bytes-per-token are known. Efficiency — what share of theoretical peak the hardware actually sustains — is not, for CPUs, because we hold zero CPU or integrated-GPU benchmarks. Our entire measured set is 14 rows, all discrete-GPU or Apple Metal.

So the ranges above are bounded by two named endpoints rather than a made-up margin:

The true figure for your machine is somewhere in between. We would rather show you that honestly than pick a midpoint and present it as knowledge.

On third-party benchmarks. We checked six published laptop measurements against physics before using any of them. Three required reading memory *faster than the memory bus allows* — one implied 99.6 GB/s on an 89.6 GB/s bus. Those were discarded. Of the three that survived, our estimate brackets each: for a 12B on single-channel DDR5 the observed figure was 4.5–5.3 tok/s, inside the range this model produces. This is the check that most published comparisons skip.

What is deliberately missing. No download sizes, no model release dates, no time-to-first-token. The first two need a model registry we could not reach from our build environment; the third needs a constant that is wrong by roughly an order of magnitude off NVIDIA hardware. Every number that is here, we checked. Full workings: our methodology and the benchmark data.

Frequently Asked Questions

What is the best local LLM for a laptop with 8GB of RAM?
A 4B-class model at Q4 — Llama 3.2 3B, Phi-4 Mini or Gemma 3 4B. All three fit 16K of context inside the roughly 6.4 GB an 8 GB laptop can give a model, with headroom to spare. Anything in the 7B–9B range either will not fit or leaves no room for context.
Can I run a local LLM without a GPU?
Yes. Generation speed is set by memory bandwidth rather than by having a GPU, so a CPU with dual-channel memory runs small models perfectly well. Expect single-digit to low-double-digit tokens per second rather than the hundreds a discrete card delivers.
Is 16GB of RAM enough to run a local LLM?
Comfortably. A 16 GB laptop gives a model about 12.8 GB, which fits a 9B at 32K context or a 14B at 16K. The 14B options are the slowest thing that fits, so a 9B is usually the better everyday choice.
How many tokens per second will I get on a laptop CPU?
Roughly 11–28 tok/s for a 2B and 3–8 tok/s for a 14B on DDR4-3200 dual-channel, rising with faster memory. These are estimates published as ranges, not measurements — we hold no CPU benchmarks of our own, so the range is deliberately wide.
Why is my local LLM so slow to start answering?
It is reading your prompt before it writes anything, and long inputs take a long time on a laptop. A pasted 4,000-word document is several thousand tokens of processing first. Shorter prompts and a smaller model both help; the delay is not a sign of a broken setup.
Do I need an NPU or a Copilot+ PC for local AI?
No. Current local runtimes largely target the CPU and GPU rather than the NPU, and text generation is bound by memory bandwidth rather than by the raw operations-per-second an NPU advertises. An NPU is good at small always-on models and a poor general-purpose LLM engine.
Should I upgrade to 32GB of RAM for local LLMs?
Only if you have already filled 16 GB and want models above 14B. More memory raises the ceiling on model size but does nothing for speed, which is set by bandwidth. Adding a second memory stick to reach dual-channel is usually the better first purchase.
Can an 8GB MacBook Air run a local LLM?
Yes, with the same 4B-class ceiling as any other 8 GB laptop. It gets about 6.0 GB for the model because macOS caps the share of unified memory available. Apple Silicon bandwidth makes it faster than a comparable x86 laptop, but it does not make it roomier.

Related

If your memory is VRAM, not system RAM: Best LLMs for 8 GB VRAM · Best LLMs for 16 GB VRAM

Set up your platform: Local AI on Windows · Local AI on macOS · Local AI on Linux

Pick by what you need it for: Privacy-sensitive work · Fully offline use · Summarising documents

How we got these numbers: Our methodology · Measured benchmarks · Check your own machine

Methodology. Estimates are labeled as estimates; verified figures link to their sources. VRAM is weights + KV cache + framework overhead; speed is a three-term latency model. Both formulas are published at /en/methodology.

← All Reports