Skip to main content
← Back to news

BenchLM publishes its September ranking of the best local LLMs: VRAM defines the tiers

Benchmarks
Source:BenchLM
The September 2026 ranking orders Open-weight models by VRAM tier: GPT-OSS 20B and ZAYA1-8B lead the 8-16 GB tier, Qwen3.6-27B (46.49) and Gemma 4 31B (44.84) dominate the 16-24 GB tier on a single RTX 4090 at Q4, and trillion-Parameter MoEs need 8x H100. A useful reference when choosing a model size for your hardware.
Read original article

Related articles

When a Big Model Isn't Needed: Choosing by Size

Memory, speed and latency: how to pick the right model size for your hardware, and when the small model genuinely wins.

Running LLM Models Locally: From First Install to First Working Model

From zero to a first answering model: what you actually need, Ollama step by step, how to pick a model for your memory, expected speeds per hardware, and the typical problems fixed.

Qwen3.8-27B vs Gemma 4 26B A4B: the new frontier at 24GB

A 27.8B dense or a MoE with 3.8B active, a 262K context against 256K and almost the same weight at Q4_K_M: two opposite ways of putting big-model quality on a 24GB consumer card. Which is faster, which goes deeper, and when to pick one over the other.

Get the next local-AI briefing

Once a week, the top selection from iaLocale: articles, news and resources on local AI, and more.

After signing up you receive a confirmation email (double opt-in). Without confirmation we send nothing. No spam, we promise.

Related resources

Tools and resources from the iaLocale archive relevant to this content.

💎

Gemma 4 26B A4B

Models

Il modello MoE della famiglia Gemma 4: 25,2B di parametri totali con 3,8B attivi per token. Contesto 256K, input testo+immagini, oltre 140 lingue, Apache 2.0. In Q4_K_M pesa circa 16,8GB: la scelta per chi ha 24GB di VRAM e vuole qualità da modello grande con velocità da modello piccolo.

🤖

OpenAI gpt-oss-20b

Models

Il modello di ragionamento open-weight di OpenAI: MoE con 21B di parametri totali e 3,6B attivi (32 esperti, top-4), contesto 128K e quantizzazione MXFP4 nativa. Licenza Apache 2.0: i pesi partono già quantizzati a 4 bit, senza passo di conversione. Sta in 16GB di VRAM e gira anche su RAM unificata o CPU.

🧠

Qwen3 30B-A3B (MoE)

Models

Mixture-of-Experts model: 30B total params but only 3.3B active per token. Small-model speed with mid-size quality; the best speed/quality ratio in its class.

🌱

Bonsai-2 27B

Models

27B reasoning model with ternary weights at 1.72 bits per weight, derived from Qwen3.8-27B with hybrid attention (~75% linear). In GGUF PTQ1_0 it takes 5.95 GB, retains 98.2% of the FP16 intelligence and reaches ~47 tok/s on an Apple M5 Max laptop, with a 262K-token context. Apache 2.0 license. Requires Prism ML's llama.cpp fork (ternary kernels are not in stock llama.cpp).

Browse all resources