Best local LLM for 32 GB

For RTX 5090 (32 GB) and Mac, 48 GB. Method

DeviceModel memory

Smartest fitQwen3.8-27BIndex 59.2 · 8-bit · 32.0 GB

Behind the frontier10 monthsFirst reached by GPT-5.2-Codex, Dec 2025

API equivalent$0.05 / 1MGLM-5.3-Flash, blended price

Models that fit in 32 GB

Top 10

#ModelIndexFormatMemoryCheapest API

01Qwen3.8-27B59.28-bit32.0 GBVultr $1.00

02Muse Glimmer 30B53.04-bit20.0 GBVultr $1.00

03Ternary Bonsai 2 27B52.28-bit31.0 GB—

04Qwen3.6-27B52.18-bit32.0 GBChutes $2.00

05Qwen3.5-27B51.38-bit32.0 GBChutes $1.75

06Qwen3.6-35B-A3B50.64-bit23.9 GBCoreWeave $0.89

07Qwen3.5-35B-A3B49.44-bit23.9 GBDeep Infra $1.00

08Gemma 4 31B47.94-bit20.9 GB—

09Gemma 4 26B A4B44.58-bit29.8 GB—

10GPT-OSS 20B41.78-bit24.3 GBCoreWeave $0.13

Index, highest first

MemorySmartest fitIndexFormatBehindAPI equivalent

8 GB Granite 4.2 8B37.64-bit23 mo$0.032

12 GB Gemma 4 12B39.04-bit23 mo$0.032

16 GB GPT-OSS 20B41.74-bit23 mo$0.032

24 GB Qwen3.8-27B59.24-bit10 mo$0.05

32 GB Qwen3.8-27B59.28-bit10 mo$0.05

48 GB Qwen3.8-27B59.28-bit10 mo$0.05

64 GB Qwen3.8-27B59.216-bit10 mo$0.05

96 GB Qwen3.8-27B59.216-bit10 mo$0.05

Models: ranked models with open weights, an official Hugging Face repository and medium or high confidence. One line per set of weights.
Memory need: the weights at the format's bits per weight (4-bit 4.85, 8-bit 8.5, 16-bit 16), plus 5% and 1 GB for a short context.
Format: the most bits that fit. The index is for the full model; a 4-bit copy can score a little lower.
Shared memory (Mac, DGX Spark, Ryzen AI Max): about two thirds goes to the model up to 36 GB, three quarters above.
Behind the frontier: months from the first release with the same index or more, among models with medium or high confidence.
Same by API: the lowest blended price (3 input : 1 output) for that index or more, per 1M tokens, among the same models.