Last updated: October 2026
Key Takeaways
- VRAM still decides what you can run. A used RTX 3090 with 24GB is still the value card at October prices; the RTX 5060 Ti 16GB is the new-card entry pick.
- The shortage got worse: the cheapest 32GB DDR5-6000 kit cost $449.97 on October 14, per Tom's Hardware, and the RTX 5090 has nearly left retail.
- Start with the computer you own, and check what your target model needs before you buy anything.
The best local AI hardware in October 2026 is a used NVIDIA RTX 3090 with 24GB of VRAM for a GPU build, an RTX 5060 Ti 16GB if you want a new card, and, above 32GB, a machine that ships with its memory installed. The memory shortage, not benchmarks, sets that answer.
This guide covers which GPU and how much VRAM to buy; for assembling an always-on box, see our local AI server build guide.
The Memory Shortage, October Update
Memory prices rose again this fall, and the graphics cards that matter for local AI followed. The cheapest 32GB DDR5-6000 kit cost $449.97 on October 14, up from $400 in August, per Tom's Hardware. Causes are in our RAM shortage explainer and our memory market analysis.
GPU asking-price floors moved the same way between August and October 11, per BuySellRam's October 13 market update: a new RTX 5090 at $4,920.99 against its $1,999 list price, a used RTX 4090 at $2,800, a used RTX 3090 at $1,350, and a new RTX 5060 Ti 16GB at $617.99 against a $429 list price.
The RTX 5090 is the extreme case. On October 10, Tom's Hardware reported a leaked factory notice saying NVIDIA had stopped making it; on October 17, it relayed NVIDIA's Jeff Fisher saying production has not been discontinued. NVIDIA's store and Micro Center no longer list it, and third-party sellers ask $7,300 to $8,500. The RTX 50 Super is reportedly on hold, and the RTX 60 series is rumored for late 2027 or 2028.
Three strategies survive this market: machines with memory already installed, used 24GB GPUs, and buying only for the models you will run.
VRAM First: The Rule That Did Not Change
VRAM capacity, not clock speed, sets the ceiling on what you can run. A model's weights must fit in video or unified memory; when they spill over, throughput drops from 30-plus tokens per second to 3 to 5.
Quantization in 2026
Quantization compresses weights from 16-bit to 4-bit, cutting memory to roughly a quarter with small quality loss; 4-bit GGUF files are the standard. Plan on about 0.6GB per billion parameters, plus 2 to 4GB for the system and context. For mixture-of-experts (MoE) models, count total parameters, not active ones.
| Model Size | Memory at 4-bit | Hardware That Holds It |
|---|---|---|
| 7B to 8B | About 5GB | Any 16GB computer, or an 8GB card |
| 12B to 14B | 7 to 9GB | 12GB to 16GB card |
| 27B to 35B | 16 to 21GB | 24GB card, or a 32GB Mac or mini PC |
| 70B dense | 40 to 42GB | Two 24GB cards, or 64GB of unified memory |
| 120B-class MoE | About 61 to 65GB | 16GB card plus 64GB of system RAM, or 96GB of VRAM |
| About 300B MoE | About 200GB (104GB at 3-bit) | 256GB machine, or 128GB at 3-bit |
| 700B and up | 365GB and up | 512GB Mac Studio or a server |
Ranges use the 0.6GB rule and our model guide's file sizes, before context. Our local AI models by VRAM guide names a pick for each tier.
Best GPU for Local LLMs (October 2026)
For most people, the best GPU for a local LLM is a used RTX 3090: 24GB at 936 GB/s, cheaper than any other 24GB card. New, the RTX 5060 Ti 16GB is the entry pick; past 32GB on one card, the step up is a 96GB workstation card or unified memory.
| GPU | VRAM | Bandwidth | Model Ceiling (Q4) | Status, October 2026 |
|---|---|---|---|---|
| RTX 3060 12GB | 12GB GDDR6 | 360 GB/s | ~13B dense | Cheapest reliable CUDA entry |
| RTX 4060 Ti 16GB | 16GB GDDR6 | 288 GB/s | ~14B dense | Now costs more new than a 5060 Ti |
| RTX 5060 Ti 16GB | 16GB GDDR7 | 448 GB/s | ~14B dense, 20B MoE | Best new entry card |
| RTX 3090 24GB | 24GB GDDR6X | 936 GB/s | ~30B class | Best value per gigabyte, used |
| RTX 4090 24GB | 24GB GDDR6X | 1,008 GB/s | ~30B class | Fastest 24GB card |
| Intel Arc Pro B70 32GB | 32GB GDDR6 ECC | 608 GB/s | ~40B class | Cheapest new 32GB card; Vulkan or SYCL setup |
| AMD Radeon AI PRO R9700 32GB | 32GB GDDR6 | 640 GB/s | ~40B class | ROCm 7.2 or Vulkan setup |
| RTX 5090 32GB | 32GB GDDR7 | 1,792 GB/s | ~40B class | Production reportedly halted |
| RTX PRO 6000 Blackwell 96GB | 96GB GDDR7 ECC | 1,792 GB/s | 120B-class MoE | Workstation ceiling, 600W |
Manufacturer specifications (B70 per StorageReview's July review); ceilings assume 4-bit with context headroom. Use the price buttons below for live listings.
The two 32GB cards are the cheapest new way past 24GB, with two catches: price and software. On October 16, Computing for Geeks put the B70 near $1,300 and the R9700 near $1,700 on Amazon, against launch prices of $949 and $1,299. Both run llama.cpp and LM Studio through Vulkan. Intel's faster SYCL path means compiling llama.cpp against its oneAPI toolkit, and StorageReview called Intel's vLLM branch "still a beta release at best." AMD's ROCm 7.2 path is easier, but still more work than CUDA.
NVIDIA stays the default because the local stack is CUDA-first.
Build Tiers, Repriced
Start With What You Own
A 16GB computer built after 2020 runs capable small models free. Install Ollama, pull a 7B to 12B model, and expect 5 to 15 tokens per second on CPU: slow but usable for summaries, drafts and document questions.
A RAM upgrade is now a price check first: buy only if today's listing still pencils out.
Check Price on Amazon: DDR4 32GB Laptop RAM Kit
Check Price on Amazon: G.SKILL Flare X5 32GB DDR5-6000 Desktop Kit
An RTX 3060 12GB is the cheapest real GPU acceleration, roughly four times CPU-only speed on small models.
Check Price on Amazon: RTX 3060 12GB
For a turnkey box instead, see our mini PC for local AI guide.
The Value Build: Used RTX 3090
The used RTX 3090 is still the value pick at October's $1,350 floor. A used RTX 4090 costs about twice as much for the same 24GB, and 24GB at 936 GB/s holds 30B-class models, including Qwen3.8-27B, entirely in VRAM.
Check Price on Amazon: RTX 3090 24GB (Renewed)
Buying new, the RTX 5060 Ti 16GB is the entry pick: 448 GB/s handles 14B dense and 20B-class MoE models at conversational speed, per Tom's Hardware's review. October's $617.99 floor is about 44 percent above its $429 list price.
Check Price on Amazon: RTX 5060 Ti 16GB
Pair a 3090 with an 850W supply, NVMe storage and fresh airflow; most used 3090s have lived hard lives. Add RAM only at a price you checked today: 64GB covers offload and video, and 96GB lets a 16GB or 24GB card run 120B-class MoE models with their experts in RAM.
Check Price on Amazon: Corsair RM850x 850W PSU
Check Price on Amazon: Samsung 990 EVO Plus 2TB NVMe
Check Price on Amazon: DDR5 64GB Desktop Kit
Check Price on Amazon: Crucial Pro 96GB DDR5 Desktop Kit (2x48GB)
Check Price on Amazon: Noctua NF-A12x25 G2 Case Fan
High End: RTX 4090, RTX 5090, Workstation Cards and Unified Memory
The RTX 4090 is the fastest 24GB card and the RTX 5090's 32GB holds 40B-class models; both sell far above list. If a 5090 listing comes from a third-party seller at $7,000 or more, walk away: two used 3090s give you 48GB for under half that.
Check Price on Amazon: RTX 4090 24GB
Check Price on Amazon: RTX 5090 32GB
Check Price on Amazon: 1000W Power Supply
The workstation ceiling is the RTX PRO 6000 Blackwell: 96GB of GDDR7 at 1,792 GB/s and up to 600W, per NVIDIA's spec sheet, enough for a 120B-class MoE model or a 70B model at 8-bit, at a professional-part price.
Check Price on Amazon: RTX PRO 6000 Blackwell 96GB
The alternative is unified memory, where CPU and GPU share one pool. Apple's M5 Ultra Mac Studio has shipped since September 22 with 1.2TB/s of bandwidth, and its 512GB configuration is "coming in late October," per Apple; our M5 Ultra wait-or-buy guide weighs each tier. Strix Halo boxes with 128GB get our Ryzen AI Max+ 395 reality check, and NVIDIA's GB10 machines (DGX Spark, ASUS Ascent GX10) are measured in our DeepSeek V4-Flash, GLM-5.3-Flash and DeepSeek V4.1-Flash reality checks. For a box that runs around the clock, see the server build guide.
Beyond Chatbots: Video, Image, and Agent Workloads
Video generation needs more memory than chat, but quantized builds fit 8GB to 16GB cards: Wan 2.2 14B drops from 54 to 65GB at full precision to about 6GB as a 480p GGUF build, and its 5B sibling runs in 8 to 12GB, per WillItRunAI's April guide. Wan 2.2 is still the newest downloadable Wan, per Alibaba's Wan-Video GitHub. Lightricks' LTX-2.5, a 22B audio-and-video model released in August, ships open weights with FP8 and offload options, free below $10 million in annual revenue.
Expect minutes per clip, and plan 64GB of system RAM, since video pipelines offload large text encoders there. Image generation runs well on 12GB to 16GB cards.
Agents need tool competence more than size: a 20B to 35B model with strong function calling on a 24GB card does real work. An agent is a server on your network, so read our secure local agent setup guide first.
Models Worth Running (October 2026)
These picks follow our model-by-VRAM guide's October check, plus gpt-oss-120b for a 16GB card with plenty of RAM.
| Memory | Pick | Local Size | Notes |
|---|---|---|---|
| 8 to 12GB | Gemma 4 E4B | 4.5GB Q4 | Apache 2.0, multimodal |
| 16GB | Gemma 4 12B Unified | 6.7GB Q4 | 256K context with room to use it |
| 24 to 32GB | Qwen3.8-27B | 16.46GB Q4 | Apache 2.0, strongest fit for a 3090 |
| 16GB card plus 64GB RAM | gpt-oss-120b | About 61GB MXFP4 | Apache 2.0 MoE, experts held in system RAM |
| 128GB | DeepSeek V4-Flash 0731 | 104GB 3-bit | MIT, mainline llama.cpp and LM Studio |
| 256GB | GLM-5.3-Flash; DeepSeek V4.1-Flash (1017) | 199.71GB Q4; 150 to 200GB resident | Both MIT; V4.1 needs fast NVMe and community builds |
| 384 to 512GB | GLM-5.3 | 365 to 467GB Q4 | Custom license; needs a 512GB machine |
| Above 512GB | Kimi K3, DeepSeek V4-Pro-0813 | 893GB and up | Downloadable, not runnable at home |
Sizes are quantized weights before context, from our VRAM guide's October 10 check, the gpt-oss-120b model files and our October 18 V4.1-Flash reality check.
The bottom row is the honest map; our Kimi K3 hardware reality check shows what a trillion-class model takes. Alibaba announced Qwen 4, including a 27B tier, on September 22, but no weights had shipped as of October 13; our Qwen 4 27B tracker follows it.
What Not to Do in 2026
- Do not buy an 8GB GPU for AI. Models and context fill 8GB immediately. Our calculator keeps an 8GB reference setup because many readers own one; it shows the least a model runs on, not what to buy.
- Do not panic-buy RAM. If you do not need the capacity this month, do not pay this market's prices for it.
- Do not wait for a rescue generation. No new consumer GeForce generation is in sight, so the card that exists today is the market.
- Do not run unverified model files. Download GGUF weights only from the Ollama library or verified publishers on Hugging Face; files from unknown uploaders are unauditable.
- Do not expose your AI box. Ollama binds to localhost by default; keep it that way, and put an always-on server on its own network segment, as our VLAN setup guide shows. For private model downloads, use Proton VPN or Mullvad, both transparently owned; we skip VPNs whose incentives conflict with yours.
Not sure what your machine runs? Our LLM VRAM calculator grades open text, image and video models against your graphics card, Mac or mini PC, in your browser, with no sign-up.
Check what your hardware can run
Frequently Asked Questions
What is the best GPU for a local LLM right now?
A used RTX 3090: in October 2026 it is the cheapest 24GB card, at 936 GB/s, and holds 27B to 32B models in VRAM at 4-bit. New, the RTX 5060 Ti 16GB is the entry pick. Above 32GB, look at a 96GB workstation card or a 128GB unified-memory machine.
Is the RTX 5090 discontinued?
Not officially. On October 10, Tom's Hardware reported a leaked factory notice saying NVIDIA had stopped production; on October 17, it relayed NVIDIA's Jeff Fisher saying production has not been discontinued. Either way, NVIDIA's store, Best Buy and Micro Center had none, and third-party listings ran about $7,300 to $8,500.
Is 2026 a bad time to buy local AI hardware?
It is the most expensive memory market in years, and waiting is not free: prices kept rising through October, and no new consumer GeForce generation is in sight. Buy only what your target models need, favor used 24GB cards and machines with memory installed, and skip speculative capacity.
How much VRAM does a 70B model need?
A 70B dense model at 4-bit needs roughly 40 to 42GB plus context headroom: two 24GB GPUs, a 48GB or larger professional card, or a 64GB unified-memory machine. A single 24GB card runs 70B only by offloading to system RAM, which cuts speed several-fold.
Can I run AI video generation locally?
Yes. Quantized Wan 2.2 builds run on 8GB to 16GB cards, and Lightricks' open LTX-2.5 adds synchronized audio. Expect minutes per clip, and plan generous system RAM, since video pipelines offload large text encoders out of VRAM.
Should I buy a Mac or an NVIDIA GPU for local AI?
NVIDIA wins on speed and software support, since the local stack is CUDA-first. Apple Silicon wins on memory per dollar, silence and power draw, and above 32GB a Mac loads models no consumer GPU can hold. Apple says the 512GB M5 Ultra Mac Studio arrives in late October.
What is the best first purchase for local AI?
Nothing, if you own a 16GB computer: install Ollama and a small model free, and learn what you use. If you are building, buy a used RTX 3090, because 24GB opens the 30B class where local AI stops feeling like a compromise.
