RTX 5070 Ti vs RTX 3090 for LLMs: 16GB new or 24GB used
RTX 5070 Ti vs RTX 3090 for local LLMs: which models fit 16 GB or 24 GB, estimated speed and context on each card, and price per tok/s at Nov 2, 2026 prices.
Updated · By RigPerDollar
As an Amazon Associate I earn from qualifying purchases. Some links on this page are affiliate links.
Short answer
For local LLMs, the RTX 5070 Ti and RTX 3090 run gpt-oss-20b and Gemma 4 12B at close estimated speeds, but only the 24 GB 3090 fits Qwen3.6-27B, Gemma 4 26B A4B and Qwen3.6-35B-A3B. At Nov 2, 2026 prices, a new 5070 Ti ($1,120) cost less than a used 3090 ($1,499) and less per estimated tok/s, while the 3090 cost less per GB.
What this guide compares
The RTX 5070 Ti is a new card with 16 GB of VRAM. The RTX 3090 is an older card with 24 GB, and our price for it is a used price. VRAM is the memory on your graphics card. The model file and its working memory both have to fit there. Context is how much text the model can hold at once, counted in tokens (pieces of words). Tok/s means tokens per second, or how fast the model writes.
Every speed here is a RigPerDollar estimate for llama.cpp (a popular open-source program for running models locally) at short context, with a target of about 15% either way. We rank by fit, context, estimated speed and estimated speed per dollar. We don’t test how good a model’s answers are.
Specs side by side
These figures come from our GPU catalog, which takes them from NVIDIA’s own spec pages and documents, such as its comparison page.
| Spec | RTX 5070 Ti | RTX 3090 |
|---|---|---|
| VRAM | 16 GB | 24 GB |
| Memory bandwidth | 896 GB/s | 936 GB/s |
| Memory type | GDDR7 | GDDR6X |
| Memory bus | 256-bit | 384-bit |
| Rated board power | 300 W | 350 W |
| Released | Feb 20, 2025 | Sep 24, 2020 |
Memory bandwidth is how fast the card can read its own memory. To write each token, the model reads its weights from memory again, so bandwidth sets most of the writing speed. The 3090 has 40 GB/s more, which is why its estimates below run a few tok/s ahead on every model. The bigger difference is the 8 GB of extra memory.
Models that fit both cards
File sizes are Q4 GGUF files, except Ternary Bonsai 2 27B, which ships as a 1.75-bit file. Q4 is a compressed format that shrinks a model so it fits in less memory. By our fit rule, a model fits when its file plus 1.5 GB of headroom fits in VRAM, and at least 4K tokens of context with an 8-bit cache still fit after that. Context numbers assume a 1 GB runtime reserve.
Each context cell shows two numbers. The first uses the default FP16 cache. The second uses an 8-bit (q8_0) KV cache. The KV cache is the model’s notes on every token it has read, and it grows with context. An 8-bit cache stores those notes in a smaller format, so more context fits.
| Model | Q4 file | Context on 16 GB (FP16 / 8-bit) | Context on 24 GB (FP16 / 8-bit) | Est. tok/s, RTX 5070 Ti | Est. tok/s, RTX 3090 |
|---|---|---|---|---|---|
| Mistral Small 3.2 24B | 14.33 GB | 3K / 7K | 51K / 97K | 52 | 54 |
| Devstral Small | 14.33 GB | 3K / 7K | 51K / 97K | 52 | 54 |
| gpt-oss-20b | 11.67 GB | full 128K / full 128K | full 128K / full 128K | 235 | 242 |
| Microsoft phi-4 | 9.05 GB | full 16K / full 16K | full 16K / full 16K | 79 | 82 |
| JetBrains Mellum2.1 12B-A2.5B Thinking | 8.07 GB | full 128K / full 128K | full 128K / full 128K | 263 | 269 |
| Google Gemma 4 12B | 7.12 GB | full 256K / full 256K | full 256K / full 256K | 98 | 102 |
| Qwen3.5-9B | 6.17 GB | full 256K / full 256K | full 256K / full 256K | 88 | 90 |
| Ternary Bonsai 2 27B | 5.95 GB | 134K / 253K | 254K / full 256K | 90 | 92 |
| K2-Horizon-7B | 5.59 GB | 62K / 117K | 115K / 217K | 121 | 125 |
| IBM Granite 4.2 8B | 5.35 GB | 57K / 108K | 105K / full 128K | 125 | 130 |
| LFM2.5-8B-A1B | 5.16 GB | full 128K / full 128K | full 128K / full 128K | 190 | 192 |
Some notes on the list:
- On the models in this table, the estimates differ by 2 to 7 tok/s. That is well inside the 15% target either way, so treat the two cards as about equal on speed for these models.
- Mistral Small 3.2 and Devstral Small pass the fit rule on 16 GB, but with only 3K of default context, or 7K with an 8-bit cache. That is enough for short chats. On 24 GB the same file gets 51K, or 97K with an 8-bit cache. Here memory decides, and speed hardly changes.
- gpt-oss-20b, Mellum2.1 and LFM2.5-8B-A1B are mixture-of-experts (MoE) models. An MoE model has many parts, and only a few of them run for each token. gpt-oss-20b has 20.91B parameters in total but only 3.6B active, which is why its estimate is so high.
- Ternary Bonsai 2 27B is PrismML’s ternary version of Qwen3.8-27B. It needs PrismML’s own fork of llama.cpp, because stock llama.cpp rejects these files.
The 16GB vs 24GB vs 32GB VRAM guide covers what each memory size runs across every card size, if you are weighing other cards too.
Models that need the 24 GB card
These fit the RTX 3090 and miss 16 GB under the fit rule. The last column shows how much more memory a 16 GB card would need.
| Model | Q4 file | Context on RTX 3090 (FP16 / 8-bit) | Est. tok/s, RTX 3090 | 16 GB short by |
|---|---|---|---|---|
| Qwen3.6-35B-A3B | 22.29 GB | 33K / 63K | 157 | 7.8 GB |
| NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB | 22.02 GB (native 2-bit) | 116K / 219K | 137 | 7.6 GB |
| MiniCPM-V 4.7 35B-A3B | 21.17 GB | 87K / 164K | 159 | 6.7 GB |
| Nanbeige4.3-33B | 20.79 GB | 10K / 19K | 38 | 6.3 GB |
| Meta Muse Glimmer 30B | 18.39 GB | full 128K / full 128K | 43 | 3.9 GB |
| Gemma 4 31B | 18.32 GB | 40K / 90K | 43 | 4.2 GB |
| GLM-4.7-Flash | 18.31 GB | 84K / 159K | 261 | 3.9 GB |
| Qwen3.6-27B | 17.98 GB | 74K / 140K | 39 | 3.5 GB |
| IBM Granite 4.2 30B | 17.72 GB | 19K / 37K | 44 | 3.3 GB |
| Qwen3.8-27B | 17.44 GB | 82K / 155K | 40 | 3 GB |
| Gemma 4 26B A4B | 17.04 GB | full 256K / full 256K | 211 | 2.6 GB |
| LFM2.5-24B-A2B | 14.53 GB | full 128K / full 128K | 179 | 0.1 GB |
Some notes on the list:
- The dense models in this group, such as Qwen3.6-27B, Qwen3.8-27B and Gemma 4 31B, estimate at about 40 tok/s on the 3090. Every part of a dense model runs for each token, so they read far more memory per token than the MoE models.
- The MoE models, such as Qwen3.6-35B-A3B, Gemma 4 26B A4B and GLM-4.7-Flash, estimate well above 100 tok/s, because only 3B to 4B parameters are active per token.
- LFM2.5-24B-A2B misses 16 GB by 0.1 GB under our rule. It is the closest miss.
- The Nemotron 3.5 Bolt row is NVIDIA’s own 24 GB build, at native 2-bit instead of Q4. NVIDIA made it to fit 24 GB of memory. Our model data notes that you should expect lower quality than a Q4 file of the same model.
- MiniCPM-V 4.7 is a vision-language model. The fit leaves out its vision file (1.12 GB), which you also need if you send images.
- K2-Horizon-32B and K2-Horizon-MoVA-36B-A4B also fit the 3090, but tightly, with 7K and 3K of default context.
- Qwen3.5-27B and Qwen3.5-35B-A3B, the earlier releases with the same file sizes as Qwen3.6-27B and Qwen3.6-35B-A3B, also fit the 3090 and miss 16 GB.
- Microsoft Fara1.5 27B fits as well. It is a computer-use agent for web browsers, so we left it out of the table.
Older models that the site now marks as superseded also need 24 GB, including Qwen3-32B and Qwen3-Coder-30B-A3B. You can push part of a model into system RAM on a 16 GB card, but that runs much slower than the numbers here. My Rig shows RAM offload fits once you enter your system RAM.
Price per tok/s and per GB
Prices are from Nov 2, 2026. The table uses gpt-oss-20b (an MoE model) and Gemma 4 12B (a dense model) as test models, because both fit every card here with full context.
| Card | Price (Nov 2, 2026) | $ per GB of VRAM | $ per est. tok/s, gpt-oss-20b | $ per est. tok/s, Gemma 4 12B |
|---|---|---|---|---|
| RTX 5070 Ti | $1,120 new | $70.00 | $4.77 | $11 |
| RTX 3090 | $1,499 used | $62.46 | $6.19 | $15 |
The RTX 5070 Ti price is the lowest in-stock Newegg listing that day, read by picksbycard. The average of the in-stock listings was $1,792.
The RTX 3090 price is used. It is the bottom of the typical range ($1,499 to $1,625) from 295 completed US eBay sales over 14 days, price plus shipping, tracked by gpusold.com. The median sale was $1,551, which works out to $64.63 per GB.
On Nov 2, 2026 prices, the used 3090 was the more expensive of the two cards. It cost $379 more than the lowest new 5070 Ti listing, and more per estimated tok/s on both test models. It still cost less per GB of VRAM, and the models that need 24 GB don’t fit on a 16 GB card under our fit rule. Running them with part of the model in system RAM is much slower. For Qwen3.6-27B, the 3090 works out to $38 per estimated tok/s.
Our 24 GB build is a full PC parts list around one used RTX 3090, with the total for the whole PC. The RTX 3090 vs RTX 5090 guide covers the step up from a 3090.
Measured runs from other people
Our calibration data holds five published single-card llama.cpp runs on these two cards. The three below are 4-bit runs of models in our tables. The other two, both on the 3090, are Qwen3.6-27B at Q6_K (31.8 tok/s), a larger file than the one we estimate, and the older Llama 2 7B.
| Card | Model and quant | Runtime | Measured tok/s | Source and date |
|---|---|---|---|---|
| RTX 3090 | Qwen3.6-27B, Q4_K_M | llama.cpp (CUDA), build b8777 | 40.14 | ahelpme.com, Jun 10, 2026 |
| RTX 3090 | Qwen3.6-35B-A3B, UD-Q4_K_M | llama.cpp (CUDA), build b10088 | 157.66 | insiderllm.com, Jul 2026 |
| RTX 5070 Ti | gpt-oss-20b, MXFP4 | llama.cpp (CUDA), build b8077 | 225.02 | llama.cpp GitHub discussion 15396, Feb 23, 2026 |
For comparison, our estimates are 39 tok/s for Qwen3.6-27B and 157 tok/s for Qwen3.6-35B-A3B on the RTX 3090, and 235 tok/s for gpt-oss-20b on the RTX 5070 Ti. These published runs are among the results our speed constants were fitted to, so close agreement is expected, and it isn’t an independent test. The 5070 Ti run also used the MXFP4 file, while our estimate uses the size of the Q4_K_M file in our model data.
Where the RTX 5060 Ti 16GB fits
The RTX 5060 Ti 16GB has the same 16 GB as the 5070 Ti, so every fit and context number in the 16 GB columns above applies to it too. It has half the bandwidth, 448 GB/s, so its estimates are lower.
| Card | Price (Nov 2, 2026) | $ per GB of VRAM | Est. tok/s, gpt-oss-20b | $ per est. tok/s | Est. tok/s, Gemma 4 12B | $ per est. tok/s |
|---|---|---|---|---|---|---|
| RTX 5060 Ti 16GB | $780 new | $48.75 | 147 | $5.31 | 52 | $15 |
The $780 price is the lowest in-stock Newegg listing that day, read by picksbycard. The average of the in-stock listings was $1,146. Per GB, the RTX 5060 Ti 16GB is the cheapest of the three. Per estimated tok/s it sits between the other two on gpt-oss-20b and at the same rounded $15 as the 3090 on Gemma 4 12B. On Mistral Small 3.2 it estimates 27 tok/s, about half the 5070 Ti’s 52. We also have a 16 GB build around this card.
Power
These are the rated board power figures from NVIDIA’s comparison page, as listed in our catalog. We have no measured power draw for any of these cards, and we don’t estimate electricity cost.
| Card | Rated board power |
|---|---|
| RTX 5060 Ti 16GB | 180 W |
| RTX 5070 Ti | 300 W |
| RTX 3090 | 350 W |
The 5070 Ti is rated 50 W below the 3090. Check that your power supply covers the card you pick, along with the rest of your PC.
Which to pick
Pick the RTX 5070 Ti if the models you want fit in 16 GB. It is new, it is rated at lower power, and on Nov 2, 2026 prices it cost less than a used 3090 and less per estimated tok/s. On the models both cards hold, the estimated speeds are close. The limit is context on the larger files: Mistral Small 3.2 and Devstral Small get only 3K of default context on 16 GB. The 5070 Ti price is a new retail listing, while the 3090 price comes from used eBay sales, so check the seller’s return terms on a used card.
Pick the RTX 3090 if you want the 27B to 35B models in the second table, such as Qwen3.6-27B, Gemma 4 26B A4B or Qwen3.6-35B-A3B, or more context on 24B models. None of those fit a 16 GB card under our rule. It also cost less per GB of VRAM.
If you want 16 GB and speed matters less than price, the RTX 5060 Ti 16GB costs less than either and fits the same models as the 5070 Ti.
To see every model’s fit, context and estimated speed on one card, open My Rig set to the RTX 5070 Ti or the RTX 3090.
What we haven’t verified
- Used prices move fast. The 3090 figure is a range from 14 days of eBay sales, and a card you find may cost more or less.
- Every speed here is an estimate from our methodology. Our reported-runs records have no runs for these three cards. The published runs above feed our calibration, so they don’t check it independently.
- We don’t test answer quality, so we can’t say which model gives better answers on either card.
- We don’t cover the condition of a used card, or warranty terms on either card.
- Power figures are rated numbers. We have not measured what these cards draw while running a model.
- Speeds are for writing at short context. Reading a long prompt and long-context slowdown are not in the estimates.