Changelog
ModelVRAM had 1 changes on October 12, 2026, the latest day below. This page lists what changed for people using the site, newest first: new calculators, model pages, data and fixes. Subscribe with the RSS feed.
- LightOnOCR-3 4B: memory per page, GGUF files and mmproj: LightOn’s new OCR model joins the calculator with its BF16 file, the image tokens and cache one page takes, the community GGUF files with the vision projector OCR needs, and LightOn’s vLLM flags.
- Kolibri-1: 78B total, 3.46B active, official FP8 files: Kolibri-1 joins the calculator with its published weight size, 40 sliding-window layers, 10 full-attention layers and the official vLLM plugin command.
- Weight memory for pplx-decider v1.1 and d1-omni: Two decision-model pages separate checkpoint memory from unmeasured working memory and show the official loaders, input limits and supported dtypes.
- Who maintains the numbers, on every page: Every page now ends with who maintains it, when its data was last checked and how far the estimates are from public measurements, and the About page sets out the method, the checks and the editorial principles.
- Mistral Large 4 VRAM, estimated before the weights: Memory for Mistral Large 4 (1.05T total, 49B active) at BF16 to 1-bit, the GPUs, Macs and RAM that hold it and its KV cache, from Mistral's figures and Mistral Large 3's files until the weights land on October 31.
- All guides on one page: Every guide on the site grouped by topic, with the question each answers and one number from it, linked from the home page, the calculator, the footer and each guide.
- Guide: big MoE models on one GPU with --n-cpu-moe and -ot: How --n-cpu-moe, --cpu-moe, -ot and --fit place a MoE model's experts, with source links, and the N, VRAM, RAM and speed for five MoE models on 16, 24 and 32 GB cards.
- Guide: why a 24 GB GPU does not give you 24 GB: What CUDA reports free on Linux and Windows in public llama.cpp logs, shared GPU memory, llama.cpp compute buffers, vLLM --gpu-memory-utilization, and how the calculator's own overhead rule compares.
- Guide: KV cache quantization, q8_0 vs q4_0 vs f16: Bytes per value of each llama.cpp cache type, how to set it in llama.cpp, vLLM and Ollama with source links, the cache of five models at 32K and 128K, and the one public perplexity comparison.
- Guide: splitting a model across two or more GPUs: How llama.cpp and vLLM divide weights and KV cache between GPUs, with source links, public logs and the memory each card needs for Qwen3.8 27B, Gemma 4 31B and Llama 3.1 70B on five multi-GPU setups.
- Guide: how much unified memory a Mac GPU can use: The default GPU limit for each Mac memory size from 16 to 256 GB with the llama.cpp logs behind it, how to raise it with iogpu.wired_limit_mb, and the largest model each size runs.
- ThinkingCap Qwen3.8 27B model page: VRAM per GGUF file and context for BottleCap AI’s shorter-thinking Qwen3.8-27B fine-tune, with the context that fits a 24 GB card and the builds that fit 16 GB.
- Accuracy page: prompt speed and image/video checks: The accuracy page now also lists all 32 llama-bench runs behind the time-to-first-token estimate, with the error per hardware family, and the ComfyUI load logs behind the image and video file sizes.
- Vendor-reported benchmark scores next to VRAM: GPQA, MMLU-Pro and LiveCodeBench scores that makers published in their own model cards, linked to the source: on model pages, as a GPQA column in the per-size lists, and as a y-axis option on the chart.
- Bespoke Nimble 9B model page: VRAM per quant and context for Bespoke Nimble 9B, with its GGUF files and Ollama tags.
- Article: the llama-server --fit context regression: Why llama.cpp builds b10999–b11200 picked contexts such as 828,160 tokens, how --fit sizes context, slots and layers now, and the flags to set.
- GPU index page: Every GPU, Mac, CPU-only PC and multi-GPU setup on one page, with how many models each runs.
- Shows cloud GPU options when a result doesn’t fit: When a model does not fit a local card, the calculators name the memory it needs and the cloud GPU class that holds it, with plain, unsponsored links.
- Load models by their Ollama name: Type an Ollama library name such as qwen3:30b-a3b into the LLM VRAM calculator to load that model.
- Apple LensVLM 9B model page: VRAM for Apple’s LensVLM-9B per quant and context, its GGUF files and mmproj, and what its visual text compression saves in KV cache.
- Latency and throughput estimates for Laya and Jev alternatives: How long one decision takes and how many per second a GPU or CPU handles, on the Laya and Jev alternatives pages.
- Ternary Bonsai 2 27B VRAM page: Real file sizes of the PTQ1_0, PQ2_0 and MLX files and the longest context on 8, 12 and 16 GB GPUs.
- llama-server --fit preview: The LLM VRAM calculator shows what llama-server’s automatic --fit would pick for context, slots and GPU layers, and the explicit flags instead.
- Chart: every model by the VRAM it needs: A scatter plot of all tracked models by VRAM at Q4_K_M against parameters, with 8–80 GB card lines.
- 2026 Local LLM VRAM Report: The largest open model per memory tier and the 32K KV cache ranking, with CSV downloads under CC BY 4.0.
- Image and video model VRAM calculator: Peak VRAM and system RAM for FLUX.2, Wan 2.2, LTX-2 and Qwen-Image 2.1 in ComfyUI, checked against public ComfyUI runs.
- Jev alternatives you can run locally: Open models that do the API-only Jev’s job on your own GPU or CPU, with file sizes and VRAM.
- Laya VRAM requirements: Real file sizes of Laya’s checkpoints in PyTorch, ggmlc and GGUF, and whether a 16 GB or 24 GB GPU or a CPU runs it.
- Predicted vs measured: The estimates next to 20 public llama.cpp and vLLM measurements, with the error of each.
- KV cache counted as llama.cpp allocates it: Gemma 4’s global layers and sliding-window caches now match llama.cpp’s logs to the byte.
- Open dataset and best LLMs per VRAM size: Every model’s VRAM estimates as CSV and JSON under CC BY 4.0, and the models that fit 8 to 96 GB.
- What can my PC run?: Detects your GPU in the browser and lists the local LLMs it runs, with the best quantization and speed.
- vLLM KV cache and concurrency calculator: The KV cache pool vLLM allocates, in tokens, and how many requests fit at once, checked against startup logs.
- Fine-tuning VRAM calculator: GPU memory for full fine-tuning, LoRA and QLoRA in Transformers or Unsloth, checked against 15 published runs.
- MoE offload calculator: The smallest llama.cpp --n-cpu-moe that fits your GPU, from real GGUF tensor sizes.
- Qwen-Image 2.1 VRAM calculator: Peak VRAM for every DiT, text encoder and VAE combination of Qwen-Image 2.1, with real measurements.
- LLM VRAM and speed calculators: The first two tools: how much GPU memory a model needs from its Hugging Face files, and how fast it writes on your hardware.
Updated .