| Hardware profile | ||||
|---|---|---|---|---|
| Hardware | Memory | Bandwidth | Computing power | vLLM |
| RTX 5090NVIDIA GeForce RTX 5090 32 GB desktop GPU class | 32 GBGPU VRAM | 1792 GB/s | ~3352 TOPS | yes |
| Compatible models Estimated speeds, not benchmark results: calculated from memory bandwidth and model size. Real results can differ significantly because there is no precise formula for deriving LLM generation speed from hardware specifications alone. | |||
|---|---|---|---|
| Model | Size | Approx. Q4 memory | Estimated generation |
| Ornith 1.0 35BMoE | 35Bactive 3B | 21 GB | 256 tok/s4-bit quantization |
| Ornith 1.5 35BMoE | 35Bactive 3B | 21 GB | 256 tok/s4-bit quantization |
| Qwen 3.6 35BMoE | 35Bactive 3B | 21 GB | 256 tok/s4-bit quantization |
| Laguna XS 2.1MoE | 33Bactive 3B | 19.8 GB | 256 tok/s4-bit quantization |
| Qwen 2.5 32BDense | 32.5B | 19.5 GB | 48 tok/s4-bit quantization |
| Gemma 4 31BDense | 31B | 18.6 GB | 50 tok/s4-bit quantization |
| Qwen 3 Coder 30BMoE | 30.5Bactive 3.3B | 18.3 GB | 242 tok/s4-bit quantization |
| Granite 4.1 30BDense | 30B | 18 GB | 52 tok/s4-bit quantization |
| Granite 4.2 30BDense | 30B | 18 GB | 52 tok/s4-bit quantization |
| Muse GlimmerDense | 30B | 18 GB | 52 tok/s4-bit quantization |
| GLM 4.7 FlashMoE | 30Bactive 3B | 18 GB | 256 tok/s4-bit quantization |
| Nemotron 3.5 LightningMoE | 30Bactive 3B | 18 GB | 256 tok/s4-bit quantization |
| Fara 1.5 27BDense | 27B | 16.2 GB | 58 tok/s4-bit quantization |
| Qwen 3.6 27BDense | 27B | 16.2 GB | 58 tok/s4-bit quantization |
| Qwen 3.8 27BDense | 27B | 16.2 GB | 58 tok/s4-bit quantization |
| Gemma 4 26BMoE | 26Bactive 4B | 15.6 GB | 213 tok/s4-bit quantization |
| LFM 2.5 24BMoE | 24Bactive 2B | 14.4 GB | 323 tok/s4-bit quantization |
| GPT-OSS 20BMoE | 20.9Bactive 3.6B | 12.5 GB | 228 tok/s4-bit quantization |
| Instella-MoE 16B ThinkMoE | 16Bactive 2.8B | 9.6 GB | 267 tok/s4-bit quantization |
| Qwen 3 14BDense | 14B | 8.4 GB | 107 tok/s4-bit quantization |
| Gemma 3 12BDense | 12B | 7.2 GB | 124 tok/s4-bit quantization |
| Gemma 4 12BDense | 12B | 7.2 GB | 124 tok/s4-bit quantization |
| Fara 1.5 9BDense | 9B | 5.4 GB | 160 tok/s4-bit quantization |
| Ornith 1.0 9BDense | 9B | 5.4 GB | 160 tok/s4-bit quantization |
| Ornith 1.5 9BDense | 9B | 5.4 GB | 160 tok/s4-bit quantization |
| Qwen 3.5 9BDense | 9B | 5.4 GB | 160 tok/s4-bit quantization |
| Qwythos 9BDense | 9B | 5.4 GB | 160 tok/s4-bit quantization |
| Granite 4.1 8BDense | 8B | 4.8 GB | 177 tok/s4-bit quantization |
| Granite 4.2 8BDense | 8B | 4.8 GB | 177 tok/s4-bit quantization |
| Qwen 3 8BDense | 8B | 4.8 GB | 177 tok/s4-bit quantization |
| LFM 2.5 8BMoE | 8Bactive 1B | 4.8 GB | 435 tok/s4-bit quantization |
| Ling 3.0 TinyMoE | 7.9Bactive 1.3B | 4.7 GB | 394 tok/s4-bit quantization |
| Fara 1.5 4BDense | 4B | 2.4 GB | 313 tok/s4-bit quantization |
| Gemma 3 4BDense | 4B | 2.4 GB | 313 tok/s4-bit quantization |
| Qwen 3 4B Instruct 2507Dense | 4B | 2.4 GB | 313 tok/s4-bit quantization |
| Qwen 3.5 4BDense | 4B | 2.4 GB | 313 tok/s4-bit quantization |
| Spark X2.5 4BDense | 4B | 2.4 GB | 313 tok/s4-bit quantization |
| Granite 4.1 3BDense | 3B | 1.8 GB | 388 tok/s4-bit quantization |
| Granite 4.2 3BDense | 3B | 1.8 GB | 388 tok/s4-bit quantization |
| Llama 3.2 3B InstructDense | 3B | 1.8 GB | 388 tok/s4-bit quantization |
| Ministral 3 3BDense | 3B | 1.8 GB | 388 tok/s4-bit quantization |
| Nanbeige 4.2 3BDense | 3B | 1.8 GB | 388 tok/s4-bit quantization |
| LFM 3 1.7BDense | 1.7B | 1 GB | 560 tok/s4-bit quantization |
Raon 1 Mini is evaluated via Performance, Memory, Reasoning, HumanEval Remix, Kanban, Sand Physics, Dungeon Crawler, Blender, and Godot tests using RTX 2000 and RTX 5090 hardware.
Tests InternScience/Sage-A1.5-Q4_K_M-GGUF on RTX 2000 and RTX 5090 using Performance, Memory, Reasoning, HumanEval Remix, Seat Planner, Sand Physics, Dungeon Crawler, Blender, and Godot benchmarks.
Tests Flint by RecursiveSI using Performance, Memory, Reasoning, HumanEval Remix, Kanban, Sand Physics, Dungeon Crawler, Blender and Godot on RTX 2000 and RTX 5090.
Demonstrates setting up AnythingLLM on Windows with Ollama to query PDFs locally, fixing the Agent mode web search issue by switching to Chat mode, and testing speed with Qwen 3.8 27B on an RTX 5090.
Tests Qwen3.8 27B Q4 on Blender and Godot, evaluates Qwen3.8 Flash Next via Strata, and compares local performance against Claude Opus 4.6 using an ASUS ROG system with RTX 5090.
Demonstrates running Claude Code locally on Windows via Ollama using models like Qwen 3.8 27B, covering installation, configuration, and usage examples such as writing expenses.py.
Tests Ming Image 0.1 Design on an RTX 5090 using FP8 conversion, generating posters, app mockups, comic panels and photo edits in about 8 seconds each.
Runs Qwen Image 2.1 locally on an RTX 5090 using fp8 and NF4 precisions, testing VRAM usage, render times across resolutions, and cost against GPT Image 2 and Nano Banana Pro.
Benchmark Qwen 3.8 27B Q4_K_M via llama.cpp on MacBook Pro M5 Max versus RTX 5090, analyzing memory usage, GGUF headers, KV cache behavior, and MTP drafting performance at 8K context.
Compares DeepSeek V4 Flash and Qwen 3.8 Flash Next using Unsloth UD-IQ1_S 1-bit builds on 19 hard problems via CPU-only and RTX 5090 expert layer offload benchmarks.
Benchmark IBM Granite 4.2, Qwen 3.8 27B, Gemma 4, and Ornith 1.5 on an RTX 5090, measuring KV cache cost, usable context, generation speed, coding, math, and tool calling via Python tests.
Benchmarking Ornith 1.5 9B and 35B, Qwen 3.8 27B, and Gemma 4 on an RTX 5090 using llama-server with 4-bit weights and FP16 KV cache via code-graded metrics.
Benchmarks Qwen 3.8 Flash Next on Intel Core Ultra 9 285K CPU versus RTX 5090 offloading using llama.cpp, measuring throughput and perplexity for 1-bit and 2-bit quantizations.
Weight-level teardown shows 131 of 866 tensors changed in Qwen 3.8 27B Uncensored via abliteration, tested on RTX 5090 with llama-server against a 30-prompt refusal benchmark.
Compares Qwen 3.8 Flash Next and Qwen 3.8 27B architectures, detailing MoE, sparse attention, gated residuals, 51B n-gram lookup tables, and RTX 5090 MTP measurements.
Compares Qwen 3.8, 3.6 and 3.5 using 397 generations on an RTX 5090, analyzing MTP layer activation, split vision encoder, post-training effects, Terminal-Bench scores and token speeds.
Tests Qwen 3.8 27B, Muse Glimmer 30B and Gemma 4 26B on an RTX 5090 using Ollama, evaluating consistency across 12 accounting questions run 10 times each.
Evaluates 45 llama.cpp configurations for Qwen 3.8 27B on an RTX 5090, analyzing speculative decoding, KV cache quantization, flash attention, and context window impacts on speed and VRAM usage.
19:16
Tests Qwen3.8-27B via NINFER on RTX 5090 with DeepSeek harness across ~300 iterations to build a 3D Acropolis scene in a single HTML/CSS/JS file.
Configures Claude Code to use Nemotron 3.5 Lightning via Ollama on an RTX 5090 through .claude/settings.local.json, verified by generating and modifying a Python palindrome function.
Side-by-side coding test using Claude Code on Ollama to generate Space Invaders, Breakout, and Tetris HTML canvas games with Qwen 3.8 27B, Muse Glimmer, and Gemma 4 on an RTX 5090.
Tests Qwen 3.8 27B, Nemotron 3.5 Lightning, and Muse Glimmer on an RTX 5090 using 16 hard math, coding, and reasoning problems via Ollama with Q4_K_M quantization, measuring accuracy, time, VRAM, and
Benchmark compares Nemotron 3.5 Lightning and Muse Glimmer 30B on an RTX 5090 using needle-in-haystack, math, logic, JSON, knowledge, and coding tests at 64,000 tokens, measuring speed, VRAM, and thin