Compare Claude and GPT LLMs with local LLMs
Key facts
Test 10 open models free in the browser, next to Claude and GPT. A used 24 GB card holds the mid-size ones. Memory and speed figures are our estimate.
Updated . Checked . watch on YouTube @localllmrig.
Test 10 open-weight local models free in the browser (no login, 0 GB of local memory), or tick them alongside Anthropic Claude and OpenAI GPT, up to 8 models at once. A single 24 GB GPU (23.4 GB usable) or a 32–36 GB Mac holds Gemma 4 31B, Qwen3.8-27B, Nanbeige4.3-33B, Recursive Flint and Qwen3.6-35B-A3B. Memory totals and speeds below are our estimate. Builder interview numbers are self-reported.
1. Watch and run the live 8-LLM trick-question test in your browser
Before downloading 15–46 GB of model weights or paying for another month of Claude or ChatGPT, run your own prompts side by side at the home chat.
- Free in the browser ($0, no account): tick up to 8 of the 10 free open models, including Gemma 4 31B, Gemma 4 26B A4B, Recursive Flint, NVIDIA Nemotron 3.5 Lightning, NVIDIA Nemotron 3.5 Super VL, and Laguna S 2.1.
- Add Claude, GPT, Gemini, and Grok from the Online only section on the home chat: Claude Sonnet 5.5, Claude Opus 5.5, Claude Haiku 5.5, GPT-6.1 Sol, GPT-6 Luna, GPT-5.6 Terra, Gemini 4 Argon, or Grok 4.7. Credits are required for those cloud models.
- In our 24-second screen recording (watch on YouTube @localllmrig), Gemma 4 31B, which fits on a Mac mini M6 32GB or a used RTX 3090 24GB, and Kimi caught a modified €1.10 bat-and-ball trick. ChatGPT and Claude missed it. Gemma 4 31B hardware.
2. Side-by-side comparison: Claude and GPT vs local open-weight LLMs (October 2026)
Every local memory figure below is for an 8,192-token conversation (about 12 pages of text) plus 0.80 GB of software working space. That total is our estimate. Licence labels are our reading of the licence, not legal advice. A 24 GB card leaves 23.4 GB usable.
| Model | Type and licence | Memory at 8,192 tokens | Hardware that holds it | Speed | Where to test |
|---|---|---|---|---|---|
| Claude Sonnet 5.5 / Opus 5.5 / Haiku 5.5 | Cloud proprietary (Anthropic) | 0 GB local (runs in Anthropic's cloud) | Any browser. $20–$200/month subscription or per-token API. Prompts are sent to the provider. | Cloud stream | Home chat, with credits |
| GPT-6.1 Sol / GPT-6 Luna / GPT-5.6 Terra | Cloud proprietary (OpenAI) | 0 GB local (runs in OpenAI's cloud) | Any browser. $20–$200/month subscription or per-token API. Prompts are sent to the provider. | Cloud stream | Home chat, with credits |
| Gemma 4 31B | Local open-weight, 31.27B dense, Apache-2.0, text and image. Parameter count is an estimate. | Q4_K_M: 18.32 GB measured file + 1.51 GB conversation memory + 0.80 GB working space = 20.63 GB (20.6 GB), our estimate | GeForce RTX 3090 24GB (€1,246 / $1,395 used) or Mac mini M6 32GB (€1,530 / $1,416, price unverified) | about 37 tokens per second on an RTX 3090, our estimate; about 39 tokens per second on an RTX 4090, our estimate; about 70 tokens per second on an RTX 5090, our estimate | Free on the home chat, no login |
| Gemma 4 26B A4B | Local open-weight, 25.2B total / 3.8B active, Apache-2.0, text and image. Parameter count is an estimate. | UD-Q4_K_M: 16.95 GB measured file + 0.38 GB conversation memory + 0.80 GB working space = 18.13 GB (18.1 GB), our estimate. UD-IQ4_XS is 14.8 GB and fits a 16 GB card (€899 / $832). | GeForce RTX 3090 24GB (€1,246 / $1,395 used) | Mixture of experts. We don't publish a speed for this design. | Free on the home chat, no login |
| Qwen3.8-27B | Local open-weight, 27.78B dense, Apache-2.0, text and image. Parameter count is an estimate. | UD-Q4_K_M: 16.46 GB measured file + 0.54 GB conversation memory + 0.80 GB working space = 17.80 GB (17.8 GB), our estimate. UD-Q6_K is 23.3 GB and still fits a 24 GB card. | GeForce RTX 3090 24GB (€1,246 / $1,395 used) | about 43 tokens per second on an RTX 3090, our estimate. Josh Gold reports about 30 tokens per second on vLLM for the Int4 file, and about 15 tokens per second of writing on llama.cpp, self-reported. Gabe Ortiz reports about 220 tokens per second with DFlash2 speculative decoding on an RTX 5090, self-reported | Check the fit page. The chat needs credits. |
| Nanbeige4.3-33B | Local open-weight, 33.12B dense coder, Apache-2.0. Our reading of the licence is that commercial use is allowed. | Q4_K_M: 20.79 GB measured file + 1.74 GB conversation memory + 0.80 GB working space = 23.33 GB (23.3 GB), our estimate. IQ4_XS is 21.2 GB and fits a 32 GB Mac. | GeForce RTX 3090 24GB (€1,246 / $1,395 used) or Mac Studio M4 Max 36GB (€1,890 / $2,116 used, price unverified) | about 32 tokens per second on an RTX 3090, our estimate; about 35 tokens per second on an RTX 4090, our estimate; about 61 tokens per second on an RTX 5090, our estimate; about 8.7 tokens per second on Mac Studio M4 Max 36GB, our estimate | Check the fit page. The chat needs credits. |
| Recursive Flint | Local open-weight, 39.79B total / 2.5B active, Apache-2.0 | UD-Q3_K_XL: 18.49 GB measured file (19.6 GB total on 24 GB). Q4_K_M: 23.39 GB measured file + 0.27 GB conversation memory + 0.80 GB working space = 24.46 GB (24.5 GB), our estimate. | 24 GB RTX 3090 at UD-Q3_K_XL (€1,246 / $1,395 used) or Mac mini M4 Pro 64GB at Q4_K_M (€1,890 / $2,116 used, price unverified) | Mixture of experts. Recursive reports more than 125 tokens per second on an M4 Pro and more than 150 tokens per second on a 64 GB RTX Spark. That is their announcement, not our measurement. | Free on the home chat, no login |
| NVIDIA Nemotron 3.5 Lightning | Local open-weight, 30B total / 3B active, OpenMDW-1.1 | Q4_0: 18.90 GB measured file + 0.05 GB conversation memory + 0.80 GB working space = 19.75 GB (19.8 GB), our estimate | GeForce RTX 3090 24GB (€1,246 / $1,395 used) or Mac mini M6 32GB (€1,530 / $1,416, price unverified) | Mixture of experts. We don't publish a speed for this design. | Free on the home chat, no login |
| Llama 3.3 70B | Local open-weight, 70.6B dense, Llama 3.3 Community License | Q4: 42.5 GB weights (measured file) + 2.7 GB conversation memory + 0.8 GB working space = 46.0 GB, our estimate | Dual RTX 3090, 64 GB RAM (46.8 GB usable, €3,592 / $3,808 used, price unverified) or Mac mini M4 Pro 64GB (48.0 GB usable, €1,890 / $2,116 used, price unverified) | about 17.7 tokens per second on Dual RTX 3090, 64 GB RAM, our estimate; about 2.9 tokens per second on Mac mini M4 Pro 64GB, our estimate, slow for chat | Check the 70B hardware guide |
| DeepSeek R1 Distill 70B | Local open-weight, 70.6B dense, MIT (distill) + Llama 3.3 Community License (base) | Q4: 45.8 GB total, our estimate. The weights are not a separate measured file on this list. | Dual RTX 3090, 64 GB RAM (46.8 GB usable) | about 17.8 tokens per second on Dual RTX 3090, 64 GB RAM, our estimate | Check the 70B hardware guide |
| NVIDIA Nemotron 3.5 Super VL | Local open-weight, 120B total / 12B active, OpenMDW-1.1, text, image and video | Q4_0: 71.80 GB measured file + 0.07 GB conversation memory + 0.80 GB working space = 72.67 GB (72.7 GB), our estimate | GMKtec EVO-X2 128GB (96.0 GB usable, €3,800 / $3,516) or Mac Studio M4 Max 128GB (€3,290 / $3,683 used, price unverified) | Mixture of experts. We don't publish a speed for this design. | Free on the home chat, no login |
| Qwen3.6-35B-A3B | Local open-weight, 35.95B total / 3B active, Apache-2.0. Parameter count is an estimate. | UD-Q4_K_XL: 22.36 GB measured file + 0.17 GB conversation memory + 0.80 GB working space = 23.33 GB (23.3 GB), our estimate | GeForce RTX 3090 24GB (€1,246 / $1,395 used) | Mixture of experts. We don't publish a speed for this design. | Check the fit page |
Open each fit page for the same figures: Gemma 4 31B, Gemma 4 26B A4B, Qwen3.8-27B, Nanbeige4.3-33B, Recursive Flint, NVIDIA Nemotron 3.5 Lightning, Llama 3.3 70B, DeepSeek R1 Distill 70B, NVIDIA Nemotron 3.5 Super VL, and Qwen3.6-35B-A3B. Hardware pages: GeForce RTX 3090 24GB, Mac mini M6 32GB, Mac Studio M4 Max 36GB, Mac mini M4 Pro 64GB, GMKtec EVO-X2 128GB, and Mac Studio M4 Max 128GB. See which machines hold Gemma 4 31B. For 70B, see how much memory a 70B model needs and a dual RTX 3090 for 70B.
3. Where local LLMs beat or match Claude and GPT, and where cloud still wins
- Privacy. Anything you send to a local llama.cpp, Ollama, vLLM, or oMLX server stays on your own machine. Cloud subscriptions and consumer web chats send prompts to remote servers unless you have an enterprise contract that says otherwise. See Local LLM for business in the EU.
- Cost after you own the hardware. A used GeForce RTX 3090 24GB is €1,246 / $1,395 used on our list (used RTX 3090 buying guide). After that, each token costs electricity, with no API fee. Light use can still be cheaper in the cloud. Use the buy or rent calculator to see how many hours pay back a card.
- Break-even against a $200 a month plan. A used GeForce RTX 3090 24GB at our list price (€1,246 / $1,395, checked 24 Oct 2026) reaches cash-flow break-even in about 7.2 months, including about $5.94 a month of electricity. That payback is our estimate. Read the used RTX 3090 break-even guide.
- Focused coding and reasoning. Dense 27B–33B models (Gemma 4 31B, Qwen3.8-27B, Nanbeige4.3-33B) and sparse models (Recursive Flint, Qwen3.6-35B-A3B, Gemma 4 26B A4B) fit on a single 24 GB GPU. See what fits in 24 GB.
- Huge multi-repository refactors and very long context. Claude Opus 5.5, Claude Sonnet 5.5, and GPT-6.1 Sol still lead there. Builders with large local machines test local models next to frontier APIs and keep routine or private work local. See Thomas Wang on a 128 GB M5 Max, Sunghoon Baek on a 128 GB DGX Spark, Josh Gold on two Arc Pro B70 cards, Gabe Ortiz on an RTX 5090, and James Meadlock on eight DGX Spark nodes. Speeds in those interviews are self-reported.
Sources and method
Checked . watch on YouTube @localllmrig
Speeds and memory are our estimate from the hardware specs, not a measured test. Prices are dated and may change. How we calculate
Questions
- How can I compare Claude and GPT LLMs with local LLMs on my own prompts?
- Open https://localllmrig.com/ to run up to 8 AI models side by side on the same prompt in your browser. 10 open-weight local models (including Gemma 4 31B, Gemma 4 26B A4B, Recursive Flint 40B, NVIDIA Nemotron 3.5 Lightning, NVIDIA Nemotron 3.5 Super VL, and Laguna S 2.1) are free to test with no login and 0 GB of local memory. You can also tick Claude (Sonnet 5.5, Opus 5.5, Haiku 5.5), GPT (GPT-6.1 Sol, GPT-6 Luna, GPT-5.6 Terra), Gemini, and Grok with credits to compare cloud and local outputs side by side before buying hardware.
- Which local LLM is best on a 24 GB GPU (RTX 3090 / RTX 4090) compared to Claude and ChatGPT?
- A 24 GB GPU leaves 23.4 GB usable by our estimate. The top open-weight models that fit a single 24 GB card at 8,192 tokens are: Gemma 4 31B at Q4_K_M (20.63 GB total estimate, about 37 tokens per second on RTX 3090, our estimate, https://localllmrig.com/can-i-run/gemma-4-31b); Qwen3.8-27B at UD-Q4_K_M (17.80 GB total estimate, or up to UD-Q6_K at 23.32 GB total, about 43 tokens per second on RTX 3090, our estimate, https://localllmrig.com/can-i-run/qwen3.8-27b); Nanbeige4.3-33B at Q4_K_M (23.33 GB total estimate, about 32 tokens per second on RTX 3090, our estimate, https://localllmrig.com/can-i-run/nanbeige-4-3-33b); Recursive Flint at UD-Q3_K_XL (19.56 GB total estimate, https://localllmrig.com/can-i-run/flint); and Qwen3.6-35B-A3B at UD-Q4_K_XL (23.33 GB total estimate, https://localllmrig.com/can-i-run/qwen3.6-35b-a3b).
- Can a Mac mini or Mac Studio run local LLMs that compete with Claude and ChatGPT?
- Yes, for daily coding, writing, summarization, and private local workflows. Mac mini M6 32GB (21.3 GB usable, €1,530 / $1,416, price unverified) holds Gemma 4 31B at Q4_K_M (20.6 GB total estimate), Qwen3.8-27B at UD-Q4_K_M (17.8 GB total estimate), NVIDIA Nemotron 3.5 Lightning at Q4_0 (19.8 GB total estimate), and Nanbeige4.3-33B at IQ4_XS (21.2 GB total estimate). Mac Studio M4 Max 36GB (27.0 GB usable, €1,890 / $2,116 used, price unverified) holds Nanbeige4.3-33B and Recursive Flint at Q4_K_M (24.5 GB). Mac mini M4 Pro 64GB (48.0 GB usable, €1,890 / $2,116 used, price unverified) also holds Llama 3.3 70B at Q4 (46.0 GB total estimate). A 36 GB Mac does not hold that 70B file.
- When should I keep using Claude or GPT in the cloud instead of a local LLM?
- Keep Claude (Opus 5.5 / Sonnet 5.5) or GPT (GPT-6.1 Sol / GPT-5.6 Terra) for multi-repository code refactors, very long context, and low-volume usage where buying a €1,246 / $1,395 used graphics card or a Mac is not justified. Use local open-weight models when prompts contain confidential code or regulated EU data, when you want no API fee per token, or when you need the machine to work offline. Many builders in https://localllmrig.com/interviews run a hybrid workflow.
Want it on paper? The GPU Buying Cheat Sheet (EU, Oct 2026) is a short PDF: graphics memory per budget and the cheapest machine per model size.