aleibovici/ollama-gpu-calculator v0.2.0

v0.2.0 (CALIB 2026.10)

on GitHub

Try it: https://aleibovici.github.io/ollama-gpu-calculator/

Calibration CALIB 2026.10. New ways to explore what your hardware can run, a model list re-verified against the Ollama library, and the project set up for open source contributions.

Added

• "What can this setup run?" table. Once a GPU is selected, every model preset is checked against it at its default quantization and sorted by fit, with required VRAM and tokens/sec. Each row can be loaded into the calculator (#24).
• Shareable links. The URL tracks the calculator inputs (?model=…&quant=…&ctx=…&gpu=rtx4090*2), opening a link restores the setup, and a Copy link button copies it (#26).
• Searchable pickers. Start typing in the model or GPU field and the list filters by name, size, vendor or capability (qwen 27b, moe, vision, rtx 40). Words match in any order (#28).
• 48 new model presets (71 in total), including Qwen 3.8 27b and 3.8-flash-next, the full Qwen 3.5 range, Nemotron 3.5 Super, Nemotron 3 Nano and Nano Omni, Llama 4 Scout and Maverick, Llama 3.3, DeepSeek-R1 8b/14b/32b/671b, DeepSeek-V3.1, Mistral Medium 3.5, Devstral Small 2, Mistral Small 3.2, Ministral 3, Granite 4.2, Gemma 3 4b/12b, Gemma 3n, Qwen3 8b/30b/235b, Qwen3-Coder 480b and Next, Qwen3-Next, Qwen3-VL 8b/30b, Flint, LFM 2.5, Ornith 1.5, Phi-4 and OLMo 3.1 (#27).
• Open source project files: CONTRIBUTING.md, CODE_OF_CONDUCT.md (Contributor Covenant 2.1), SECURITY.md, this changelog, issue forms, a pull request template, CODEOWNERS, .editorconfig and .gitattributes (#14).
• CI that lints, tests and builds every pull request; ESLint with React Hooks rules; Dependabot for Bun dependencies and GitHub Actions (#14, #21).

Changed

• All 71 model presets re-verified against ollama.com/library. Parameter counts, context limits and vision projector sizes now come from Ollama's tag metadata (#27).
• The deploy workflow installs from the frozen lockfile and runs the tests before publishing (#14).
• README rewritten with features, how the estimates work, link format, project structure and a privacy note (#14).
• Dependencies: Vitest 5, ESLint 10, updated Radix UI, Testing Library and React lint plugins, and actions/checkout v7 (#15, #22, #23).
• package.json now has license, repository and description metadata; the unused Create React App eslintConfig was removed (#14, #21).

Removed

• Ollama Cloud presets (qwen3.5:cloud, gpt-oss:120b-cloud) and the cloud-only readout. Cloud models run on Ollama's servers, so they don't need a local GPU (#29).

Fixed

• Mixture-of-experts models listed as dense (qwen3-coder 30b, qwen3.6 35b), which made their tokens/sec far too low (#27).
• Presets for models or sizes that don't exist on Ollama: minimax-m2 (removed), glm-5 9b (replaced by glm-4.7-flash), qwen3.5 32b (replaced by 27b and 35b) and granite4 8b (replaced by Granite 4.2) (#25, #27).
• VRAM underestimated for gemma4 e2b/e4b, which store 4.65B/7.52B parameters rather than their 2B/4B "effective" size (#27).
• Context limits: qwen3 32b is 40K on Ollama, not 128K. qwen3.6 now accepts images (#27).
• muse-glimmer 30b no longer shows the MLX-only warning on GPUs that use the GGUF build (#27).

Full changelog: v0.1.0...v0.2.0