Last updated: October 2026
Key Takeaways
- One verified pick per memory tier, from an 8 GB card to a 512 GB Mac Studio, with every file size sourced and every license checked in October 2026.
- Three picks changed: Gemma 4 E4B takes 8 to 12 GB, Qwen3.8-27B at 8-bit takes 48 to 96 GB, and GLM-5.3-Flash takes 256 GB after entering mainline llama.cpp on September 30.
- Total parameters set your memory bill; active parameters only set speed. At 4-bit, plan on about 0.6 GB per billion parameters, plus room for context.
The best local AI model for your machine depends on one number: how much memory the model can live in. As of October 2026, run Gemma 4 E4B on 8 GB, Gemma 4 12B on 16 GB, Qwen3.8-27B from 24 GB through 96 GB, DeepSeek V4-Flash 0731 on 128 GB, GLM-5.3-Flash on 256 GB, and GLM-5.3 on 384 GB and up. If you have not picked hardware yet, start with our local AI hardware guide.
Choosing a model first?
For the full ranking of open-weight models, their licenses, and how they compare to Claude and GPT, see our best open-source LLMs guide.
Jump to your tier: 8 to 12 GB, 16 GB, 24 to 32 GB, 48 to 96 GB, 128 to 256 GB, 384 GB and up, or the vetting checklist.
How to Read the Tiers
Find the memory your model can use, then read that row. Total parameters set the memory bill, because every expert in a mixture-of-experts (MoE) model must be resident; the active count only sets speed. The quantized file size is your floor, before the operating system and context. Our TurboQuant explainer covers compression.
| Tier | Typical hardware | What fits |
|---|---|---|
| 8 to 12 GB | RTX 3060 12 GB, older GPUs, 8 to 16 GB laptops | Up to about 14B dense at 4-bit; small MoE |
| 16 GB | RTX 4060 Ti 16 GB, most modern laptops | Up to about 20B dense with context room |
| 24 to 32 GB | RTX 3090, 4090, or 5090; 32 GB mini PCs and Macs | 27B to 35B dense at 4-bit; mid-size MoE |
| 48 to 96 GB | 64 GB mini PCs, dual GPUs, 64 to 96 GB Macs | 27B to 35B at 8-bit with long context; 70B dense at 4-bit |
| 128 to 256 GB | Ryzen AI Max+ 395 boxes, M5 Max and M5 Ultra Macs, DGX Spark and other GB10 machines | Up to about 300B MoE at 3-bit on 128 GB; at 4-bit on 256 GB |
| 384 to 512 GB | 512 GB Mac Studio, multi-GPU rigs, servers | 400B to 750B frontier-class MoE at 4-bit |
Leave 10 to 15 percent of memory free. On Windows, Strix Halo machines give the GPU at most 96 GB of their 128 GB.
8 to 12 GB: Gemma 4 E4B
The pick is Google's Gemma 4 E4B: Apache 2.0, multimodal, and 4.5 GB at Q4 by Google's own loading table, leaving context room on an 8 GB card. It replaces LFM2.5-8B-A1B because Artificial Analysis scores it 9 to LFM2.5's 7, and it carries no revenue clause.
LFM2.5-8B-A1B stays as the speed alternate: 8.3B total, 1.5B active, 5.16 GB at Q4_K_M, and free commercial use only under $10 million in revenue. On a 12 GB card, the 16 GB pick also fits. Watch entry: PrismML's Ternary-Bonsai-2-27B squeezes Qwen3.8-27B into about 6 to 7 GB, but only PrismML's own llama.cpp fork loads it.
Check Price on Amazon: RTX 3060 12GB
16 GB: Gemma 4 12B Unified
The pick is unchanged: Gemma 4 12B Unified, dense, Apache 2.0, with a 256K context and 6.7 GB at Q4 per Google's memory table. It reads text, images, and audio. Gemma 4 26B A4B fits the same tier at 14.4 GB and scores 17 to the 12B's 14 on Artificial Analysis.
For coding, JetBrains' Mellum2 (12B total, 2.5B active, Apache 2.0) ships official 8.1 GB Q4_K_M files; Ollama loads it from JetBrains' Hugging Face repository rather than its own library.
Check Price on Amazon: RTX 4060 Ti 16GB
24 to 32 GB: Qwen3.8-27B
The pick is Qwen3.8-27B, and independent scoring now backs it: Artificial Analysis rates it 34, the highest of anything that fits this tier. It is dense, 27.78B parameters, Apache 2.0, with image and video input and a 262K context. Unsloth's UD-Q4_K_M file is 16.46 GB; Ollama's qwen3.8:27b tag is 18 GB. Our Qwen3.8-27B hardware guide has per-machine speeds.
Alibaba announced Qwen 4, including a 27B tier, on September 22, but no weights exist as of October 10; our Qwen 4 27B tracker follows it. Two alternates each win on one axis: Nvidia's Nemotron 3.5 Lightning (30B total, 3B active, 18.9 GB at Q4_0) for latency-bound agent loops, and Meta's Muse Glimmer 30B (Apache 2.0, 16.76 GB at Q4_K_M, with a bundled draft model for 3.1x decode speed) for tool use.
The hardware anchors are the renewed RTX 3090, still the value king of local AI, and the 32 GB mini PC for an always-on box; our mini PC guide covers setup.
Check Price on Amazon: RTX 3090 24GB (Renewed)
Check Price on Amazon: Beelink SER9 PRO+ (32GB)
48 to 96 GB: Qwen3.8-27B at Full Quality
The pick is Qwen3.8-27B again, at higher quality: its 8-bit Q8_0 file is 29.05 GB, leaving a 64 GB machine room for the full 262K context. No larger model that fits under 96 GB and carries an independent score rates higher, so spend the headroom on precision and context. Gemma 4 31B (17.5 GB at Q4) is the smoother-running alternative.
For agent work, Nex-N2.5-mini (September 8) replaces Nex-N2-mini: a 35B MoE under Apache 2.0 with image input and community GGUF files at 22.3 GB for Q4_K_M. Its scores come from the lab's own harness. Watch entry: Aleph Alpha's Kolibri-1 (78B total, 3.46B active, Apache 2.0) runs from 48 GB in community builds but needs a patched llama.cpp; see our Kolibri-1 reality check.
The hardware anchor is the 64 GB unified-memory mini PC. With DRAM at multi-year highs, buy machines with the memory installed.
Check Price on Amazon: MINISFORUM X1 Pro 370 (64GB)
128 to 256 GB: Two Picks, Split by Memory
This tier now splits at the memory line: on 128 GB the pick is DeepSeek V4-Flash 0731, and on 256 GB it is GLM-5.3-Flash.
128 GB: DeepSeek V4-Flash 0731, With Qwen3.8-Flash-Next Beside It
DeepSeek V4-Flash 0731 is 284B total with 13B active, MIT-licensed, and the best-documented model at this size: Unsloth's 3-bit file is 104 GB, backed by a measured quality ladder. Mainline llama.cpp and LM Studio run it; Ollama retired its library page on August 27. Our V4-Flash reality check has the tier map.
Qwen3.8-Flash-Next scores higher, 40 to 34 on Artificial Analysis, and fits at 4-bit in 111 GB, with image input, mainline llama.cpp support since August 27, and a local 120 GB Ollama tag. It ranks second on license, not quality: the Qwen Community License 1.0 requires a separate license for model-as-a-service businesses and attribution at large scale. Home use is unaffected.
256 GB: GLM-5.3-Flash
GLM-5.3-Flash became the pick when mainline llama.cpp merged its support on September 30 (build b11279). It is 320B total with 18B active, MIT-licensed, reads images and video, and scores 42 on Artificial Analysis, the highest in this tier. Its 4-bit UD-Q4_K_XL file is 199.71 GB. On a 128 GB Linux machine it loads at 1-bit to 3-bit at about 15 to 18 tokens per second, a capable preview; our GLM-5.3-Flash reality check lists every quant. Ollama offers it only as a cloud model.
The alternate is Xiaomi's MiMo-V2.6-Flash: 309B total, 15B active, MIT, with image, video, and audio input, mainline llama.cpp support, and a 167 GB MXFP4 file; it scores 38. September's pick, Step 3.7 Flash, keeps its Apache 2.0 license, but its estimated 19 now trails both.
This tier runs on Apple's M5 Max and M5 Ultra Macs, Ryzen AI Max+ 395 boxes, and NVIDIA's GB10 machines; our M5 wait-or-buy breakdown and Ryzen AI Max+ 395 reality check run the math. Confirm 128 GB in any Strix Halo listing title; the same chassis ships with 64 GB.
Check Price on Amazon: GMKtec EVO-X2 (128GB)
Check Price on Amazon: NVIDIA DGX Spark (128GB)
Watch: DeepSeek V4.1-Flash
V4-Flash's MIT-licensed successor is a 552B backbone plus 196B of Engram lookup memory kept on an SSD. Two linked 128 GB DGX Sparks run it at 25 to 41 tokens per second in community recipes; mainline llama.cpp does not support it yet. Our V4.1-Flash reality check tracks each tier.
384 to 512 GB: GLM-5.3
The pick is GLM-5.3: about 744B total with 40B active, a 1M-token context, and an Artificial Analysis score of 45, second among open models. Its weights use a custom GLM-5.3 License, not MIT, with a security-review clause for the largest model-as-a-service operators; our GLM-5.3 open-weights ledger has the details. Mainline llama.cpp runs it, and Unsloth's 4-bit files measure 365 to 467 GB, so a 512 GB machine holds it.
The Apache-licensed alternate is now Nex-N2.5-Pro: 397B total, image input, mainline llama.cpp support, and 220 GB at IQ4_XS, with self-reported scores. Xiaomi's MiMo-V2.6-Pro holds the top open score, 46, but fits 512 GB only at 2-bit (419 GB).
Above this tier sit models you can license but not house: Kimi K3 at 2.8 trillion parameters, Qwen3.8-2.4T-A95B, Nex-N2.5-Max, and DeepSeek V4-Pro-0813, tracked in our DeepSeek V4 ledger. Downloading is free; the silicon and the electricity are not.
Fix it, don't live with it
WiFi acting up? The right gear fixes it.
Tested refurbished gear. 90-day warranty and 30-day returns.
What's your WiFi problem?
How to Vet a Model Before You Run It
Check the license, the evidence, and the tooling before you download. Weights are passive data files; the real risks are capability claims that were never real and custom loader code you never read.
- Confirm the license. A fine-tune inherits its base's terms, and one lab can ship two licenses: Qwen3.8-27B is Apache 2.0, Qwen3.8-Flash-Next is not.
- Check who ran the benchmarks. Vendor numbers are marketing until reproduced, which is why this page weighs Artificial Analysis over model cards.
- Check tooling. Files that load in mainline llama.cpp, Ollama, or LM Studio mean low friction; a fork or bundled loader makes you the QA department.
- Treat silence as data. If a model claims frontier wins and no third party confirms after weeks, wait.
Once a model passes, our zero-cost local agent stack guide covers turning it into something useful.
All Picks at a Glance
| Model | Params (total / active) | License | Local size | Best tier |
|---|---|---|---|---|
| Gemma 4 E4B | 4B effective | Apache 2.0 | 4.5 GB Q4 (Google) | 8 to 12 GB |
| LFM2.5-8B-A1B | 8.3B / 1.5B | LFM v1.0 (revenue clause) | 5.16 GB Q4 (measured) | 8 to 12 GB (speed alternate) |
| Gemma 4 12B Unified | 12B dense | Apache 2.0 | 6.7 GB Q4 (Google) | 16 GB |
| Qwen3.8-27B | 27.78B dense | Apache 2.0 | 16.46 GB Q4, 29.05 GB Q8 (measured) | 24 to 96 GB |
| Nemotron 3.5 Lightning | 30B / 3B | OpenMDW-1.1 | 18.9 GB Q4_0 (measured) | 24 to 32 GB (alternate) |
| Nex-N2.5-mini | 35B MoE | Apache 2.0 | 22.3 GB Q4 (community) | 48 to 96 GB (agents) |
| DeepSeek V4-Flash 0731 | 284B / 13B | MIT | 104 GB 3-bit (measured) | 128 GB |
| Qwen3.8-Flash-Next | 125B / 6B, plus a 51B n-gram table | Qwen Community 1.0 | 111 GB Q4 (measured) | 128 GB (alternate) |
| GLM-5.3-Flash | 320B / 18B | MIT | 199.71 GB Q4 (measured) | 256 GB |
| MiMo-V2.6-Flash | 309B / 15B | MIT | 167 GB MXFP4 (measured) | 256 GB (alternate) |
| GLM-5.3 | About 744B / 40B | GLM-5.3 License (custom) | 365 to 467 GB Q4 (measured) | 384 to 512 GB |
| Nex-N2.5-Pro | 397B MoE | Apache 2.0 | 220 GB IQ4_XS (community) | 384 to 512 GB (alternate) |
"Measured" sizes come from official, ggml-org, or Unsloth files; "community" means a third-party quantizer. Add several GB for context. Scores cited on this page are the Artificial Analysis Intelligence Index v4.3.2, read October 5, 2026. Picks re-verified October 10, 2026.
Frequently Asked Questions
What is the best local AI model for 16 GB of RAM or VRAM?
Gemma 4 12B Unified: dense, 256K context, Apache 2.0, and 6.7 GB at Q4 by Google's own loading table, which leaves real context room on a 16 GB machine. Gemma 4 26B A4B fits the same tier at 14.4 GB with a higher independent score but less headroom. The stronger Qwen3.8-27B needs a 24 GB card.
What is the best model for a 24 GB GPU in 2026?
Qwen3.8-27B: dense, Apache 2.0, image and video input, a 262K context, and 16.46 GB at 4-bit. Artificial Analysis scores it 34, at least double either 30B-class alternative. Alibaba has announced Qwen 4, including a 27B tier, but as of October 10, 2026 no weights exist.
What is the best local AI model for a 128 GB machine?
DeepSeek V4-Flash 0731: MIT-licensed, 104 GB at 3-bit, and supported by mainline llama.cpp and LM Studio. Qwen3.8-Flash-Next scores higher and fits at 4-bit in 111 GB, but its custom license restricts some commercial services. GLM-5.3-Flash loads on 128 GB only at 1-bit to 3-bit.
Can I run DeepSeek V4 at home?
V4-Flash, yes: the 0731 weights run on a 128 GB machine at 3-bit through mainline llama.cpp. V4-Pro-0813 is about 893 GB as downloaded and needs data-center hardware. V4.1-Flash runs on two linked DGX Sparks through community recipes, not in mainline llama.cpp.
Should I wait for Qwen 4 or cheaper RAM before trying local AI?
No. Start with the memory you have. Models improve faster than memory gets cheaper, so your current machine runs better AI every quarter: three of this page's tier picks changed between September and October alone. Upgrade when a specific model you want demands it, not in anticipation.
