Which LLM on RTX 4070 Ti Super (16 GB) ?

#RTX 4070 Ti Great for a local LLM: what 16 GB can do

A RTX 4070 Ti Super runs 4B to 24B models in Q4 with ease. It has 16 GB of GDDR6X on a 256-bit bus, or 672 GB/s, and 8,448 CUDA cores according to NVIDIA. This combination puts it above all 8 and 12 GB cards: Gemma 4 12B fits with plenty of room for context, gpt-oss 20B and Mistral Small 24B, each 14 GB, fit in VRAM with about 2 GB to spare, and a 5 to 8 GB model runs close to the limit of its bandwidth. What does not fit is the next tier: Qwen 3.5 27B weighs in at 17 GB and exceeds the card's capacity. The 4070 Ti Super is therefore not a card for 30B models, but one of the most balanced options for everything below that.

Architecture

Ada Lovelace, AD103 chip, the same as the RTX 4080 but with fewer active units.

Memory

16 GB of GDDR6X, a 256-bit bus, and 672 GB/s of bandwidth according to Wikipedia.

Compute

8,448 CUDA cores according to the NVIDIA product page, 4th-generation Tensor Cores.

Power

285 W of total graphics power; NVIDIA specifies a 700 W minimum power supply for the entire PC.

#4070, 4070 Super, 4070 Ti, 4070 Ti Super: which one for an LLM

All four cards share the same product-line name but aren't equivalent for LLMs. Three of them are limited to 12 GB of VRAM; only the Ti Super moves up to 16 GB and a 256-bit bus. For local AI, this difference matters more than raw compute power: it determines which models the card can handle.

The RTX 4070 Ti lists 504 GB/s according to Wikipedia, compared with 672 GB/s for the Ti Super. If you were looking for a standard RTX 4070, the dedicated page covers the 12 GB variant. One thing they all have in common: 12B models fit, while those with 14 GB or more require 16 GB.

• Which LLM on RTX 4070, 4070 Super, and 4070 Ti (12 GB)

#Which models fit in 16 GB

The 14 and 15 GB models leave little room for the context cache: Ollama uses 4 096 tokens by default, which works, but 16 000 tokens require quantizing the K/V cache. Devstral Small 2, a coding model, is right at the limit: at 15 GB, almost nothing remains for context. For coding, Mistral Small 24B or a smaller model with a longer context is a better choice on 16 GB.

#What remains out of reach for 16 GB

Three model families exceed the card’s capacity. The 27B to 35B models in Q4: Qwen 3.5 27B weighs 17 GB and Qwen 3.5 35B weighs 24 GB according to Ollama, while Gemma 4 26B ranges from 16 to 19 GB. Then there are the 70B models, whose weights alone are around 40 GB according to the site’s reference point. Finally, long-context models pushed to their maximum: an 8 GB model advertised with 256,000 tokens of context will never fit with a full cache. For these cases, move to 24 GB of VRAM, a Mac with unified memory, or a dedicated machine: the hardware page compares these options.

• Which PC for local AI

#Install and verify VRAM placement

1. 01
   Driver

   Ollama requires NVIDIA driver 550 or later for recent GeForce cards. Check your version with nvidia-smi.
2. 02
   Installation

   Install Ollama for your system, without a separate CUDA installation.
3. 03
   First test

   Launch ollama run mistral-small for a first model that uses the 16 GB, then ollama ps.
4. 04
   Control

   The PROCESSOR column should show 100% GPU; otherwise, reduce the context.

export OLLAMA_FLASH_ATTENTION=1
export OLLAMA_KV_CACHE_TYPE=q8_0
export OLLAMA_CONTEXT_LENGTH=16384
ollama serve

The K/V cache in q8_0 uses about half the memory of f16, the default setting, and requires Flash Attention. This setting is what allows a 14 GB model to serve a context of 16 000 tokens on 16 GB.

#What speed to expect: ceilings and third-party measurements

No throughput is presented here as an in-house measurement. The generation ceiling is bandwidth divided by weight size. A public third-party benchmark (LLaMA 3 8B in Q4_K_M under llama.cpp, May 2024) records 106 tokens/s on a RTX 4080 with 716.8 GB/s of bandwidth, or about 68% of its ceiling, and 75% on a RTX 4070 Ti. The 4070 Ti Super’s bandwidth is close to that of the 4080, so its order of magnitude is too. The table applies 70% of the ceiling.

gpt-oss 20B is a mixture-of-experts model: it reads only part of its weights per token, so its throughput exceeds that of a dense 14 GB model. No sourced figure is reproduced here. The main takeaway is that the card goes from fast on an 8B to adequate on a 24B.

#4070 Ti Super vs. the 4080 Super and 5070 Ti

All three cards offer 16 GB: capacity is identical, so the list of models is too. Only generation speed changes, in proportion to bandwidth. The 4080 Super gains about 10%; the 5070 Ti, about one-third. On the used market, the price difference matters more than these throughput differences: check the price tracker to compare. A used 4070 Ti Super at a good price remains an excellent choice; a new 5070 Ti is justified by its warranty and speed.

• Which LLM on RTX 4080 and 4080 Super
• Which LLM on RTX 5070 Ti
• Local AI graphics card prices

#Use cases that benefit from 16 GB

The 16 GB mainly changes what you can run at the same time. A local RAG combines a generation model, an embedding model, and a context window long enough to hold the retrieved excerpts: with Gemma 4 12B (7.7 to 8 GB) and a small embedding model, everything fits with several gigabytes to spare. A coding assistant needs only an 8B to 12B model and a 16,000-token context, with no constraints. A 14 GB model such as Mistral Small 24B takes up almost the entire card: it becomes an exclusive choice, not one model among several.

A potential pitfall for shared use: in Ollama, multiple parallel requests on the same model increase the size of the reserved context accordingly. Serving three colleagues with a 14 GB model can therefore overflow the card, while a single user works without difficulty. On 16 GB, limit parallelism or choose a lighter model for multi-user access.

#Buying a used 4070 Ti Super: what to check

The card was released in January 2024: most used examples are less than three years old, but nothing guarantees how they were used before. Prices change every week: check the price tracker rather than relying on a fixed figure in this guide. Before buying, verify that nvidia-smi shows 16 GB of memory, run a 14 GB model, monitor the temperature during a long generation, and listen to the fans. Ask for the invoice and the manufacturer's remaining warranty: they are often transferred with the card.

#2026 verdict

Keep it if

Your models fit in 16 GB. There is no reason to replace it before you need 24 GB.

Buy it used if

The price is significantly lower than that of a new 5070 Ti. Check the remaining warranty, temperature, and fan noise.

Aim for 24 GB if

You want Qwen 3.5 27B or more: no setting will fit 17 GB into 16 GB with context, and a 24 GB card saves you from buying again.

• Source: official NVIDIA specifications (RTX 4070 family)
• Source: Ollama's official FAQ (Flash Attention, K/V cache)
• Source: Mistral Small size in the Ollama library

#Frequently asked questions