Local LLMs on a gaming laptop: laptop GPUs, VRAM and speed

On a gaming laptop, look up the GPU's VRAM before its name: a laptop chip can carry a desktop card's name with less memory and less bandwidth. The RTX 4090 Laptop has 16GB, where the desktop RTX 4090 has 24GB. At an 8k context, Llama 3.1 8B Runs great on all seven laptop GPUs in CanRun's data. Qwen3 14B fits fully from 12GB up, with little margin at 12GB, gpt-oss-20b from 16GB, and a 27B dense model such as Qwen3.8 27B only on the 24GB RTX 5090 Laptop.

Laptop GPUs in CanRun, by VRAM

CanRun's catalog has seven NVIDIA laptop GPUs, and they come in four memory sizes:

• 8GB: the GeForce RTX 4060 Laptop and the GeForce RTX 4070 Laptop
• 12GB: the GeForce RTX 4080 Laptop and the GeForce RTX 5070 Ti Laptop
• 16GB: the GeForce RTX 4090 Laptop and the GeForce RTX 5080 Laptop
• 24GB: the GeForce RTX 5090 Laptop

NVIDIA's comparison page also lists the RTX 5060 and RTX 5050 Laptop with 8GB and the RTX 4050 Laptop with 6GB, which CanRun's catalog does not have yet. For what fits, an 8GB chip matches the 8GB pair above and a 6GB chip fits less; speed depends on each chip's own bandwidth.

The table counts every model in the catalog by verdict on each chip, at an 8k context with 32GB of system RAM and Windows, and names the largest model that fits on the GPU alone. Models join the catalog often, so use the counts to compare chips, not as exact figures.

In CanRun's calculation, fit depends only on the memory a chip can use, so chips of the same size fit the same models. CanRun assumes Windows with the GPU also drawing the desktop, which keeps a little VRAM back: an 8GB laptop chip can use 7.4 GB, and the 24GB RTX 5090 Laptop 23.4 GB.

Many gaming laptops do not work that way. In NVIDIA's original Optimus design, which hybrid mode still uses, the integrated graphics draws the screen and the GeForce chip only renders; Advanced Optimus and a GPU-only setting let the GeForce chip drive the screen itself. If your laptop runs in hybrid mode, untick "This GPU drives my display" in the calculator, which leaves the chip a little more usable memory. Keep it ticked when the laptop is set to use only the GeForce GPU.

Once a model fits, its generation speed follows memory bandwidth, how fast the chip reads its own VRAM, and that is where chips of the same size differ. For why, see Tokens per second for local LLMs: what's fast enough and what sets the speed. The RTX 4080 Laptop reads 432 GB/s and the RTX 5070 Ti Laptop 672 GB/s. Qwen3 14B fits on both, but it Runs well on the RTX 4080 Laptop and Runs great on the RTX 5070 Ti Laptop. Qwen3 14B can think before it answers, which costs many more tokens per answer, so CanRun holds it to speed thresholds 1.5× higher than the legend at the end shows. The RTX 4080 Laptop lands just under that line, inside its error band, while the RTX 5070 Ti Laptop clears it with room to spare.

Some chips show the opposite: different names, same results. NVIDIA lists the RTX 4060 Laptop and the RTX 4070 Laptop with the same 8GB on a 128-bit bus, and CanRun's data gives both 256 GB/s. Their generation estimates match, so they get the same verdicts: Llama 3.1 8B Runs great on both and Qwen3 14B Runs slowly on both. The RTX 4070 Laptop has more compute, which raises CanRun's prompt-speed estimate but not its generation estimate. The RTX 5080 Laptop and the RTX 5090 Laptop share their bandwidth too, 896 GB/s. A model that fits on both gets the same estimate on both; the extra 8GB on the RTX 5090 Laptop changes what fits, not how fast it runs.

Same name, different card: laptop and desktop compared

The grid pairs each laptop GPU with the desktop card that shares its name. Every cell assumes an 8k context, an f16 KV cache (the memory that holds the context), 32GB of DDR5 system RAM and Windows. Each cell shows a verdict, an estimated speed band and a confidence label, and cells labelled theoretical get a note under the grid.

Two pairs share their memory size: the RTX 4060 Laptop with the RTX 4060 (8GB), and the RTX 5080 Laptop with the RTX 5080 (16GB). The other five laptop chips have 4GB or 8GB less VRAM than the desktop card. Every laptop chip in the grid also has less memory bandwidth than its namesake.

8GB: RTX 4060 Laptop and RTX 4070 Laptop

The RTX 4060 Laptop is the closest match. It has the same 8GB, with bandwidth a little under the desktop card's (256 GB/s against 272 GB/s), so the same models fit and the estimates differ only slightly. gpt-oss-20b, a MoE model that keeps part of its expert weights in system RAM on an 8GB card, Runs well on both. Because gpt-oss-20b always thinks, the line it has to clear for that is 1.5× the one in the legend, and both the desktop RTX 4060 and the laptop chip clear it. Both are theoretical estimates (±30%) a few percent apart, so treat the two as equal.

The RTX 4070 Laptop is where the name misleads most. The desktop RTX 4070 has 12GB; the laptop chip has 8GB and about half the bandwidth (256 GB/s against 504 GB/s). Qwen3 14B at Q4_K_M, a 4-bit quant, needs 10.5 GB at 8k. It fits on the desktop card and Runs great there, but that is more than the 7.4 GB the laptop chip can use, so part of it runs from system RAM and it Runs slowly. Gemma 3 12B, outside the grid, splits the same way.

12GB: RTX 4080 Laptop and RTX 5070 Ti Laptop

Both 12GB laptop chips carry the name of a 16GB desktop card, and gpt-oss-20b is the model that separates 12GB from 16GB: it needs 12.3 GB at 8k. On the desktop RTX 4080 and RTX 5070 Ti it fits fully and Runs great; the RTX 5070 Ti cell is a public benchmark result rather than an estimate. On the two laptop chips, the expert weights that do not fit go to system RAM, and it Runs well as a theoretical estimate. Its model card says it runs within 16GB of memory; in CanRun's calculation a 16GB laptop GPU holds it fully and a 12GB one does not.

Qwen3 14B fits fully on both 12GB chips at 8k, with less than a gigabyte to spare, so check ollama ps. At 16k with an f16 KV cache it spills into system RAM, and a q8_0 (8-bit) KV cache brings it back onto the GPU. Since Qwen3 14B can think at length, 8k is a short context for it; the trade-off is explained in KV cache explained: how context length eats VRAM.

16GB: RTX 4090 Laptop and RTX 5080 Laptop

The RTX 4090 Laptop has 16GB against the desktop card's 24GB, and a little over half its bandwidth (576 GB/s against 1,008 GB/s). gpt-oss-20b fits fully and Runs great. The dense Qwen3.8 27B needs 17.6 GB, so part of it goes to system RAM and it Runs slowly on the laptop chip, while the desktop RTX 4090 runs it fully on the GPU and Runs great. Qwen3.5 35B-A3B fits fully on the desktop card and Runs great; on the laptop chip some of its experts stay in system RAM and it Runs well, a theoretical estimate.

The RTX 5080 Laptop and the desktop RTX 5080 both have 16GB, and their bandwidth is close (896 GB/s against 960 GB/s). For the five models in the grid they get the same verdicts, and the Qwen3.5 35B-A3B cells on both are theoretical estimates.

24GB: RTX 5090 Laptop

The RTX 5090 Laptop has 24GB, as much as a desktop RTX 4090, not the 32GB of the desktop RTX 5090. For the five models in the grid it still gets the same verdicts as the desktop RTX 5090.

The desktop card's extra 8GB shows up in two other places. The first is models that need more than 24GB but fit in 32GB. Nemotron 3.5 Lightning 30B-A3B needs 25.1 GB at 8k, more than the 23.4 GB the laptop chip can use. On the RTX 5090 Laptop part of its experts go to system RAM and it Runs well; on the RTX 5090 it fits fully and Runs great. Both are theoretical estimates. The second is context. The table shows Qwen3.8 27B on the RTX 5090 Laptop at each context length and KV cache type:

With an f16 KV cache, 64k needs 21.9 GB and still fits on the GPU. At 128k it needs 26.8 GB, more than the chip can use, so part of the model moves to system RAM. A q8_0 KV cache brings 128k back onto the GPU with little margin, and a q4_0 (4-bit) cache leaves more room. On the desktop RTX 5090, 128k with an f16 cache stays on the GPU. The table gives each cell's verdict.

Power limits and heat: what CanRun does not calculate

The same laptop GPU does not get the same power in every laptop. NVIDIA's laptop comparison page gives each chip a range of GPU subsystem power instead of one figure. The range is 35 W to 115 W for the RTX 4060 Laptop and the RTX 4070 Laptop, 80 W to 150 W for the RTX 4090 Laptop and 95 W to 150 W for the RTX 5090 Laptop. Two laptops with the same GPU name can therefore run it at very different power. If the maker lists the GPU power in its spec sheet, note it when you compare laptops. On a laptop you already own, NVIDIA Control Panel > System Information lists the GPU's Maximum Graphics Power; treat it as the configured limit, not a measured draw.

CanRun's calculation leaves power and heat out. Its speed estimate for a laptop chip uses the bandwidth on the spec sheet and the same efficiency factor as desktop NVIDIA cards. CanRun's data has no public benchmark of a laptop GPU yet. llama.cpp's public scoreboard for NVIDIA cards includes a few laptop results, but it does not record the power they ran at. How much speed a lower power limit or a hot chip costs is something CanRun has not measured, so read a laptop's speed band with more caution than a desktop card's.

Fit is a different matter. VRAM does not change with the power limit, so the memory results in this guide, which models fit and where the context runs out, hold for any laptop with that chip.

To learn what your own laptop does, measure it, plugged in and on battery if you use it both ways. How to read the generation rate in Ollama, llama.cpp and LM Studio is covered in Tokens per second for local LLMs: what's fast enough and what sets the speed.

First make sure the model runs on the GeForce chip. In Ollama's server log, the inference compute line should name the GeForce GPU with library=CUDA. Ollama turns Vulkan on by default, and its documentation warns that a system with both integrated and discrete graphics can pick the integrated GPU; setting GGML_VK_VISIBLE_DEVICES to the discrete GPU's index fixes that. For a model that CanRun shows fully on the GPU, ollama ps should read 100% GPU in the PROCESSOR column. An MoE model with experts in system RAM shows a CPU/GPU split there by design. If something looks wrong, the steps are in Ollama or LM Studio not using your GPU? Check the split and fix the offload.

System RAM and MoE models

A mixture-of-experts (MoE) model splits most of each layer into expert blocks, and a router picks a few of them for each token. Each token reads only a small part of the weights, but the whole file has to be in memory. When it does not fit in VRAM, llama.cpp can keep the shared weights and as many experts as fit on the GPU and leave the rest in system RAM (--n-cpu-moe). CanRun calculates the same split, so on a laptop, system RAM becomes part of the answer. The mechanics are covered in Running MoE models bigger than your GPU: gpt-oss-120b with system RAM.

Qwen3.5 35B-A3B uses about 3B of its parameters per token. At Q4_K_M it needs 22.3 GB, about three times what the RTX 4060 Laptop can use, so most of its experts sit in system RAM:

With 32GB of RAM it Runs well on this chip in CanRun's calculation, a theoretical estimate. With 16GB of RAM the Q4_K_M file does not fit. The smaller IQ4_XS file does, and Runs well, also a theoretical estimate. The 12GB RTX 4080 Laptop and RTX 5070 Ti Laptop also need 32GB of RAM for the Q4_K_M file. A 16GB chip holds more of the experts in VRAM, so on the RTX 4090 Laptop and the RTX 5080 Laptop 16GB of RAM is enough. On the RTX 5090 Laptop the whole model fits on the GPU and Runs great.

For gpt-oss-120b, too little RAM means it does not fit at all. The file is 63.4 GB. In CanRun's calculation, the 12GB to 24GB laptop chips need 64GB of RAM and the 8GB chips need 96GB. With that RAM in place it still Runs slowly on the RTX 5090 Laptop, while the desktop RTX 5090 with the same 64GB Runs well. Both are theoretical estimates.

Before you plan on more RAM, check the maker's spec sheet for the most RAM the laptop takes and whether its RAM can be upgraded at all. RAM speed matters as well, because the experts in RAM are read at RAM speed. CanRun assumes DDR5-5600 in dual channel (90 GB/s), and slower RAM lowers the estimate.

If you are considering a MacBook or a Strix Halo laptop

A MacBook's GPU has no VRAM of its own. It uses unified memory shared with the rest of the system, and CanRun counts about 70% of it as usable by default. Apple's MacBook Pro spec page lists the M5, M5 Pro and M5 Max, with up to 128GB on the M5 Max with the 40-core GPU (checked October 2026). The table counts CanRun's verdicts for the M5 Pro and the 40-core M5 Max at each memory size in its data:

Between the RTX 5090 Laptop and the 40-core M5 Max, memory size and bandwidth trade places. The RTX 5090 Laptop reads 896 GB/s and the 40-core M5 Max 614 GB/s. For a model that fits in 24GB, CanRun's estimate favours the RTX 5090 Laptop: Qwen3.8 27B Runs great on it and Runs well on a 40-core M5 Max with 48GB.

For a model far larger than 24GB, such as gpt-oss-120b, memory size wins. With 128GB, the 40-core M5 Max holds gpt-oss-120b fully in GPU memory and Runs great. The RTX 5090 Laptop needs 64GB of system RAM for it and then Runs slowly, a theoretical estimate. A model only a little over 24GB is a closer call. Nemotron 3.5 Lightning 30B-A3B keeps most of its experts on the RTX 5090 Laptop's GPU, and CanRun's estimate there is about the same as on a 40-core M5 Max with 48GB. The Mac still gets the better verdict, Runs great against Runs well, because Runs great needs the whole model on the GPU. Both are theoretical estimates.

Against the cheaper gaming laptops, a Mac can win on both counts. The M5 Pro reads 307 GB/s, more than the 256 GB/s of the RTX 4060 Laptop, and with 24GB it holds gpt-oss-20b fully in GPU memory and Runs great. The RTX 4060 Laptop's 8GB pushes part of that model into system RAM, so it Runs well at best, and only as a theoretical estimate. Which Mac configuration fits which model is covered in Running local LLMs on a Mac (M4, M5): how much unified memory?

Windows laptops built on AMD's Ryzen AI Max+ 395 (Strix Halo) also share one pool of memory between the processor and its graphics, and AMD's product page lists up to 128GB. With 128GB, gpt-oss-120b fits in GPU memory and Runs great; that result is a public benchmark, not an estimate. With 64GB it does not fit. The memory is soldered, so you choose its size when you buy, and its bandwidth, 256 GB/s, is no higher than the 8GB GeForce laptop chips'. CanRun's data does not separate laptops from small desktops with the same chip, and a laptop may run it at a lower power limit, so read its speed with the same caution as the GeForce laptop chips. What fits at each memory size is covered in Strix Halo vs DGX Spark vs Mac Studio for local LLMs: what 128 GB gets you.

Where to go next

• How much VRAM do local LLMs need? GPU requirements for 8–32 GB (2026): the memory formula and what fits on desktop cards from 8GB to 32GB.
• Running MoE models bigger than your GPU: gpt-oss-120b with system RAM: how much system RAM MoE models need, card by card.
• Tokens per second for local LLMs: what's fast enough and what sets the speed: what sets the speed and how to measure your own.
• Running local LLMs on a Mac (M4, M5): how much unified memory?: the same questions for unified memory.
• Strix Halo vs DGX Spark vs Mac Studio for local LLMs: what 128 GB gets you: 128GB machines, including Strix Halo.
• GPU pages such as the GeForce RTX 4060 Laptop and the GeForce RTX 5090 Laptop list every model in the catalog.

How sure are these numbers?

The memory figures are calculated by formula from real GGUF file sizes, each file's header and each model's published configuration, so they carry no band. The laptop chips' VRAM matches NVIDIA's specifications. It does not depend on the power limit, so the fit results apply to every laptop with that chip.

The speeds are estimates with a band and a label. Measured means a public benchmark for that card, model and quant; CanRun has none for laptop GPUs yet, so the laptop cells in this guide are estimates. Calibrated means ±12% for dense models fully on an NVIDIA or AMD card and ±20% for MoE models, partial offload and other hardware. Theoretical means ±30%, for setups whose real speeds vary widely, such as MoE experts in system RAM. The bands for laptop chips assume the chip reaches its spec-sheet bandwidth and do not account for the power limit or heat, as the power section explains.

All verdicts assume Windows with the GPU also drawing the desktop (untick that option in the calculator if your laptop's integrated graphics draws it), 32GB of DDR5-5600 dual-channel RAM, an 8k context, an f16 KV cache, one model loaded and one request at a time. Ollama and llama.cpp make their own memory estimates, so a model that only just fits here can end up partly in system RAM. ollama ps shows what happened on your laptop.

Models with a thinking mode face speed thresholds 1.5× higher, which the legend does not show. In this guide that covers gpt-oss-20b and gpt-oss-120b, which always think, and Qwen3 14B, Qwen3.8 27B, Qwen3.5 35B-A3B and Nemotron 3.5 Lightning 30B-A3B, which can switch thinking on and off. EXAONE 4.0 32B has thinking off by default but is held to the same thresholds. Llama 3.1 8B and Gemma 3 12B have no thinking mode.

FAQ

Is a laptop GPU slower than the desktop card with the same name?

In CanRun's data, yes. Each of the seven laptop GPUs has less memory bandwidth than its desktop namesake, and five of them also have less VRAM, so a model can fit on the desktop card and spill into system RAM on the laptop chip. The power limit, which CanRun does not model and which differs from laptop to laptop, can lower a laptop's speed further.

How much VRAM should a laptop for local LLMs have?

Start from the models you want to run. 8GB holds 8B-class models such as Llama 3.1 8B. 12GB holds Qwen3 14B at 8k, with little margin, and 16GB holds gpt-oss-20b fully on the GPU. A 27B to 32B dense model such as Qwen3.8 27B or EXAONE 4.0 32B needs the 24GB of the RTX 5090 Laptop to fit fully at 8k. MoE models can go further with their experts in system RAM.

Does system RAM matter on a gaming laptop?

For MoE models, yes. Qwen3.5 35B-A3B at Q4_K_M does not fit on an RTX 4060 Laptop with 16GB of RAM and fits with 32GB, with most experts in RAM. gpt-oss-120b needs 64GB of RAM on the 12GB to 24GB laptop chips and 96GB on the 8GB ones. For a dense model that does not fit in VRAM, RAM only holds the overflow, and the model slows down sharply. Before buying, check whether the laptop's RAM is soldered.

Gaming laptop or MacBook: which handles local LLMs better?

It depends on the size of the models. For a model that fits in 24GB, CanRun estimates the RTX 5090 Laptop faster than a 40-core M5 Max, because it reads its memory faster. For a model far larger than 24GB, memory size decides: gpt-oss-120b fits fully in GPU memory on a 40-core M5 Max with 128GB and on a Strix Halo laptop with 128GB, while the RTX 5090 Laptop needs 64GB of system RAM before it fits at all. The section on MacBooks and Strix Halo laptops compares the speeds.

Sources

• NVIDIA: compare GeForce laptop GPUs: memory size, memory interface width and GPU subsystem power range for each RTX 50 and RTX 40 laptop GPU.
• Apple: MacBook Pro tech specs: the M5, M5 Pro and M5 Max chips, their unified memory options and memory bandwidth.
• llama.cpp: performance of llama.cpp on NVIDIA GPUs: public llama-bench results on NVIDIA cards, including a few laptop GPUs listed without their power limit.
• llama.cpp: server README: --n-cpu-moe N, which keeps the MoE weights of the first N layers on the CPU.
• Ollama FAQ: the PROCESSOR column of ollama ps.
• Ollama docs: hardware support: Vulkan enabled by default, and GGML_VK_VISIBLE_DEVICES for systems with both integrated and discrete graphics.
• NVIDIA: Advanced Optimus: in the original Optimus design the integrated graphics drives the display; Advanced Optimus lets the GeForce GPU drive it directly.
• Razer support: maximum graphics power: where NVIDIA Control Panel shows the GPU's Maximum Graphics Power.
• AMD Ryzen AI Max+ 395: laptop processor with Radeon 8060S graphics and up to 128GB of memory.
• gpt-oss-20b model card: total and active parameters, runs within 16GB of memory.
• Qwen3.5-35B-A3B model card: total and active parameters.
• Qwen3-14B model card: switching between thinking and non-thinking modes.