All five cards have the same 16 GB of memory, so they all run the same models. GPT OSS 20B fits with its whole context window, but the strongest open model that fits at all, Qwen3.8 27B, only just fits at 3-bit. Pricing varies from $590 to $2,394 in our data, and bandwidth varies threefold. We would choose the RTX 5070 Ti. By our estimate it generates text at 93% of an RTX 5080's speed, for less than half the 5080's current street price. If you run Linux, the RX 9070 is cheaper, and at today's prices the 5080 is hard to justify for local models.
Key takeaways
- A coding assistant has enough room on 16 GB. The best open model that fits needs 3-bit, with about 27K tokens of context left for an agent working on a few files.
- The main difference is memory bandwidth. By our estimate, the fastest card on dense models generates about three times as fast as the slowest.
- The RTX 5070 Ti generates text at about 93% of a 5080's speed by our estimate, for half the street price. Its cost per token is the lowest of the cards with a street price in our data.
- The memory shortage of 2026 has affected this size in particular. Nvidia's 24 GB versions of the 5070 Ti and 5080 are on hold, leaving 16 GB for this generation's mid-range.
Five cards, one memory size
Nvidia and AMD released all five between January and June 2025, with list prices ranging from $349 for the RX 9060 XT 16GB to $999 for the RTX 5080. On frame rates, the reviews were mixed. KitGuru put the 5080 only 9 to 13% ahead of the 4080 Super it replaced, in games and at the same price, calling it "hardly going to set the world alight." The $749 RTX 5070 Ti "delivers almost the same power for $250 less", wrote The Shortcut. Club386 recommended the RX 9070, though it told readers it was "worth saving up" for the RX 9070 XT, then $50 more.
For local models, the shared memory is more important than gaming performance. The weights have to be in GPU memory, and all of them are read for every generated token. The amount of memory determines which models you can use, and the bandwidth determines how quickly they run. The order of these cards depends on memory bandwidth.
In 2026, these cards suffered from the memory shortage. In January, Hardware Unboxed reported that ASUS had placed the RTX 5070 Ti in end-of-life status, with the 5060 Ti 16GB close to the same fate. Club386 explained that 16 GB cards "require twice as many memory chips as the 8GB models". The next day ASUS said the cards were "neither discontinued, nor described as end-of-life (EOL)" and were only short of supply. The 24 GB "Super" versions of the 5070 Ti and 5080 had been expected early in 2026. In July, Notebookcheck described the cards as "finished and ready for launch", held back because their 3 GB memory chips were still "too expensive". Nvidia has never announced them.
What 16 GB holds in late 2026
16 GB is more than enough for a coding assistant. Memory is tight for agents using the best open models, with only 3-bit fitting and the context size limited by what remains after the weights. An agent soon fills that space with file contents and tool output.
| Model | Longest context on 16 GB | Best quant on 16 GB | Memory at 16K context | llmrun Score |
|---|---|---|---|---|
| GPT OSS 20B | full 131K window | Q4_K_M | 13.1 GB | 96.4 |
| Qwen3.5 9B | full 262K window | Q4_K_M | 6.6 GB | 99.4 |
| Qwen3 14B | full 41K window | Q4_K_M | 11.9 GB | 98.0 |
| Qwen3.6 35B A3B | about 41K tokens | IQ3_S | 15.7 GB | 106.8 |
| Gemma 4 26B A4B | about 31K tokens | Q4_K_S | 15.5 GB | 105.8 |
| Qwen3.8 27B | about 27K tokens | Q3_K_L | 15.4 GB | 121.9 |
The first three models leave room to spare. Only GPT OSS 20B of these three has a coding score, with its full 131K window fitting on all five cards. It is fast for its size because only about 4 billion parameters out of 21 billion need to work on each token (why that matters).
The other three have higher scores, with room for 27K to 41K tokens of context. This gives sufficient room for an agent session over a few files, although a whole repository is too much. Qwen3.8 27B has by far the highest score, and only this model has an agents score here. On 16 GB you trade quant for context: a smaller 3-bit IQ3_S quant gives about 71K, with the steps in our Qwen3.8 27B memory guide. For the 4-bit Q4_K_M you need 17.9 GB and a 24 GB card such as a used RTX 3090. A 70B model cannot fit at any quant: Llama 3.3 70B needs 51.3 GB at Q4_K_M.
Where the five pull apart
The estimated order from fastest to slowest is the RTX 5080, 5070 Ti close behind, RX 9070, RTX 5060 Ti 16GB and RX 9060 XT 16GB, on every model.
| Model | RTX 5080 | RTX 5070 Ti | RX 9070 | RTX 5060 Ti 16GB | RX 9060 XT 16GB |
|---|---|---|---|---|---|
| Qwen3.5 9B | S, Runs greatQ4_K_M · ~98 tok/s | S, Runs greatQ4_K_M · ~92 tok/s | A, Runs wellQ4_K_M · ~60 tok/s | A, Runs wellQ4_K_M · ~46 tok/s | A, Runs wellQ4_K_M · ~30 tok/s |
| Qwen3 14B | S, Runs greatQ4_K_M · ~66 tok/s | A, Runs wellQ4_K_M · ~61 tok/s | A, Runs wellQ4_K_M · ~40 tok/s | A, Runs wellQ4_K_M · ~31 tok/s | B, DecentQ4_K_M · ~20 tok/s |
| Qwen3.8 27B | B, DecentQ3_K_L · ~42 tok/s | B, DecentQ3_K_L · ~39 tok/s | C, UsableQ3_K_L · ~26 tok/s | C, UsableQ3_K_L · ~19 tok/s | D, Barely runsQ3_K_L · ~13 tok/s |
| GPT OSS 20B | S, Runs greatQ4_K_M · ~183 tok/s | S, Runs greatQ4_K_M · ~173 tok/s | S, Runs greatQ4_K_M · ~124 tok/s | S, Runs greatQ4_K_M · ~98 tok/s | A, Runs wellQ4_K_M · ~68 tok/s |
| Gemma 4 26B A4B IT | B, DecentIQ4_NL · ~173 tok/s | B, DecentIQ4_NL · ~164 tok/s | B, DecentIQ4_NL · ~119 tok/s | B, DecentIQ4_NL · ~95 tok/s | B, DecentIQ4_NL · ~66 tok/s |
| Qwen3.6 35B A3B | C, UsableIQ3_S · ~181 tok/s | C, UsableIQ3_S · ~174 tok/s | C, UsableIQ3_S · ~132 tok/s | C, UsableIQ3_S · ~108 tok/s | C, UsableIQ3_S · ~77 tok/s |
- S, Runs greatGPT OSS 20BQ4_K_M · ~183 tok/s
- S, Runs greatQwen3.5 9BQ4_K_M · ~98 tok/s
- S, Runs greatQwen3 14BQ4_K_M · ~66 tok/s
- B, DecentGemma 4 26B A4B ITIQ4_NL · ~173 tok/s
- B, DecentQwen3.8 27BQ3_K_L · ~42 tok/s
- C, UsableQwen3.6 35B A3BIQ3_S · ~181 tok/s
- S, Runs greatGPT OSS 20BQ4_K_M · ~173 tok/s
- S, Runs greatQwen3.5 9BQ4_K_M · ~92 tok/s
- A, Runs wellQwen3 14BQ4_K_M · ~61 tok/s
- B, DecentGemma 4 26B A4B ITIQ4_NL · ~164 tok/s
- B, DecentQwen3.8 27BQ3_K_L · ~39 tok/s
- C, UsableQwen3.6 35B A3BIQ3_S · ~174 tok/s
- S, Runs greatGPT OSS 20BQ4_K_M · ~124 tok/s
- A, Runs wellQwen3.5 9BQ4_K_M · ~60 tok/s
- A, Runs wellQwen3 14BQ4_K_M · ~40 tok/s
- B, DecentGemma 4 26B A4B ITIQ4_NL · ~119 tok/s
- C, UsableQwen3.6 35B A3BIQ3_S · ~132 tok/s
- C, UsableQwen3.8 27BQ3_K_L · ~26 tok/s
- S, Runs greatGPT OSS 20BQ4_K_M · ~98 tok/s
- A, Runs wellQwen3.5 9BQ4_K_M · ~46 tok/s
- A, Runs wellQwen3 14BQ4_K_M · ~31 tok/s
- B, DecentGemma 4 26B A4B ITIQ4_NL · ~95 tok/s
- C, UsableQwen3.6 35B A3BIQ3_S · ~108 tok/s
- C, UsableQwen3.8 27BQ3_K_L · ~19 tok/s
- A, Runs wellGPT OSS 20BQ4_K_M · ~68 tok/s
- A, Runs wellQwen3.5 9BQ4_K_M · ~30 tok/s
- B, DecentGemma 4 26B A4B ITIQ4_NL · ~66 tok/s
- B, DecentQwen3 14BQ4_K_M · ~20 tok/s
- C, UsableQwen3.6 35B A3BIQ3_S · ~77 tok/s
- D, Barely runsQwen3.8 27BQ3_K_L · ~13 tok/s
By our estimate, every card is faster than a reader on the dense Qwen3.8 27B, the most demanding model here. The average adult reads non-fiction at about 238 words a minute (roughly four words a second), and our estimate for the slowest card on the 27B is about 13 tokens a second. Our estimates for mixture-of-experts models have a smaller spread of about 2.4 times, because each token reads only a few gigabytes of weights.
Waiting is where the budget cards cost you. The card must read all the thousands of tokens in the agent's prompt on every turn before writing begins.
- RTX 50804.5
- RX 9070 XT5.3
- RTX 5070 Ti5.8
- RX 90707.1
- RX 9060 XT 16GB10
- RTX 5060 Ti 16GB10.7
| Item | seconds to first token, estimated, Qwen3.8 27B Q3_K_L |
|---|---|
| RTX 5080 | 4.5 |
| RX 9070 XT | 5.3 |
| RTX 5070 Ti | 5.8 |
| RX 9070 | 7.1 |
| RX 9060 XT 16GB | 10 |
| RTX 5060 Ti 16GB | 10.7 |
Ten seconds is fine once in a while, but a coding agent sends a long prompt on every turn. For AMD, these estimates are probably optimistic. For the record, the 5070 Ti's speed reading a 7B prompt is 1.3 to 1.5 times as fast as the 9070 XT in llama.cpp's public scoreboards (CUDA, Vulkan), and the XT probably would not beat it in practice.
Results obtained from Hardware Corner's Qwen3 14B measurements on the 5060 Ti, 5070 Ti and 5080 at a 16K context were within 7% of our generation estimates. Its short-prompt runs were faster still.
RTX 5060 Ti 16GB and RX 9060 XT, the cheap seats
There are smaller 8 GB versions of both budget cards under nearly the same name, with both 5060 Ti sizes on NVIDIA's spec page. None of the larger models above can fit in 8 GB, so make sure the size on the box is right before you pay.
The RTX 5060 Ti 16GB is the least expensive new card with 16 GB for CUDA. At its $429 list price, it has a 180 W rating and GDDR7 memory on a narrow 128-bit bus with 448 GB/s of memory bandwidth. We use the list price for comparison because we have no reliable street price, and it applies only if a card is available near that price. The RX 9060 XT 16GB has older GDDR6 memory at 320 GB/s and the lowest power draw of the five at 160 W. AMD's suggested price was $349. In our October data a new one cost $590.
- RTX 5060 Ti 16GB
- RX 9060 XT 16GB
Qwen3.5 9B
RTX 5060 Ti 16GB45.8RX 9060 XT 16GB30.2Qwen3.8 27B (Q3_K_L)
RTX 5060 Ti 16GB19.4RX 9060 XT 16GB12.8GPT OSS 20B (MoE)
RTX 5060 Ti 16GB97.9RX 9060 XT 16GB67.5Qwen3.6 35B A3B (MoE)
RTX 5060 Ti 16GB107.5RX 9060 XT 16GB77.1
| Item | RTX 5060 Ti 16GB | RX 9060 XT 16GB |
|---|---|---|
| Qwen3.5 9B | 45.8 | 30.2 |
| Qwen3.8 27B (Q3_K_L) | 19.4 | 12.8 |
| GPT OSS 20B (MoE) | 97.9 | 67.5 |
| Qwen3.6 35B A3B (MoE) | 107.5 | 77.1 |
Our estimate for the AMD card is probably low. The 9060 XT generates at roughly three-quarters of the 5060 Ti's speed on a 7B model in llama.cpp's scoreboards, and is closer than its bandwidth would suggest. Either way, both run GPT OSS 20B comfortably and the 27B slowly.
We would choose the 5060 Ti 16GB if it is available near list price. Another concern is the 9060 XT at $590, only about $200 below an RX 9070 with twice its bandwidth. Choice of the 9060 XT is only justified with a hard limit on power or budget.
The RX 9070 is AMD's best case, with a Linux catch
The RX 9070 has a 256-bit GDDR6 bus with 640 GB/s of bandwidth and a power rating of 220 W. By our estimate it runs at about two-thirds of the 5070 Ti's speed on every model, and llama.cpp's scoreboards give the same ratio on a 7B. In our October data it cost $800 new against $1,170 for the 5070 Ti, with about the same cost per token as the NVIDIA card and 80 W less power.
The RX 9070 XT has the same memory, so it generates text at the same estimated speed. The extra compute helps with prompts: by our estimate it starts answering a long prompt to the 27B almost two seconds sooner, and on the Vulkan scoreboard it reads a 7B prompt 1.6 to 2 times as fast as the 9070. The cost is higher power consumption, at 304 W against 220. Our October averages put the XT about $160 higher, and we would choose the plain 9070 for chat and the XT only for long-prompt agent work.
That is the catch. You need to consider the operating system with AMD. CUDA on NVIDIA works the same on Windows and Linux. On a supported Linux release that you already run, the RX 9070 represents the best value here.
The RTX 5070 Ti is the 16 GB card we'd pick
The RTX 5070 Ti uses the same GB203 chip as the RTX 5080 with fewer cores, and GDDR7 bandwidth of 896 GB/s against the 5080's 960. For generating text, nearly everything depends on bandwidth, and by our estimate it gives 93 to 96% of the 5080's speed on every model here. A Founders Edition was never available, and at launch Notebookcheck's review called it "only lucrative if you can get one at the Nvidia-mandated $749 MSRP".
Nobody gets it at $749 now. ASUS briefly called the 5070 Ti end-of-life, and of the four cards we track month by month, its price has risen fastest this year.
- New, Oct 2026
- $1,339
- New, Jan 2026
- $998
- Launch price (MSRP)
- $749
| Month | New | Used |
|---|---|---|
| Jan 2026 | $998 | — |
| Feb 2026 | $1,062 | — |
| Mar 2026 | $1,082 | — |
| Apr 2026 | $1,115 | — |
| May 2026 | $1,143 | — |
| Jun 2026 | $1,152 | — |
| Jul 2026 | $1,167 | — |
| Aug 2026 | $1,210 | — |
| Sep 2026 | $1,293 | — |
| Oct 2026 | $1,339 | — |
We use our 13 October street price ($1,170, against October's $1,339 median) in the comparison here. Even at that median price, it works with every runner on Windows and Linux, and our estimates have it running 27B models at 3-bit and every smaller model quickly for much less than the card above.
The RTX 5080 costs twice as much for a few percent
The RTX 5080 is the fastest card here, and it is the hardest to recommend for local models. By our estimate it is about 7% faster than the 5070 Ti at generating text. On llama.cpp's CUDA scoreboard the two are within 3% on a 7B. Both have the same 16 GB of memory. Our latest street price was $2,394, from eight listings on 14 November, and October's median was $1,928. The 5080 costs about twice as much as a 5070 Ti on single-day prices.
| Hardware | Price | Speed | Power W | $ per 1M tokens |
|---|---|---|---|---|
| RTX 5080 | $2,394street · Nov 14 | 98 (best) | 360 | $1.70 |
| RTX 5070 Ti | $1,470street · Nov 21 | 92 | 300 | $1.15 |
| RX 9070 XT | $815street · Oct 13 | 60 | 304 | $1.07 |
| RX 9070 | $800street · Oct 13 | 60 | 220 | $1.00 |
| RX 9060 XT 16GB | $590street · Oct 13 | 30 | 160 (best) | $1.47 |
| RTX 5060 Ti 16GB | $429MSRP (best) | 46 | 180 | $0.75 |
- RTX 5080$2,394street
- RTX 5070 Ti$1,470street
- RX 9070 XT$815street
- RX 9070$800street
- RX 9060 XT 16GB$590street
- RTX 5060 Ti 16GB$429MSRP
- USD · lower is better
- RTX 5080360
- RTX 5070 Ti300
- RX 9070 XT304
- RX 9070220
- W · lower is better
- RTX 5080$1.70
- RTX 5070 Ti$1.15
- RX 9070 XT$1.07
- RX 9070$1.00
- RX 9060 XT 16GB$1.47
- RTX 5060 Ti 16GB$0.75
- USD per 1M tokens · lower is better
The table spreads each card's price plus three years of electricity at four hours a day over its estimated output on a 9B model. The 5060 Ti looks cheapest per token only because the table uses its $429 list price.
For the price of a 5080 you can look outside this group. A used RTX 3090 cost about $1,500 in our October data and has 24 GB of memory, 8 GB more. It runs Qwen3.8 27B at the 4-bit Q4_K_M quant at about 35 tok/s by our estimate, close to a 5080 on the same model at 3-bit. If you already have a 5080 or also game at 4K, it runs everything here well. As a purchase for local models alone, it is the weakest choice here.
Verdict
Our choice
We would choose the RTX 5070 Ti, since the 5080 adds only a few percent for nearly twice the money. The RX 9070 represents about the same value with less power on a Linux release that ROCm supports.
Good for
- The RTX 5070 Ti for CUDA with no OS restrictions, at about 93% of a 5080's estimated speed
- The RX 9070 if you run a Linux release that ROCm supports: the same cost per token and 80 W less power
- The RTX 5060 Ti 16GB as the entry card, if you can find it near its $429 list price
Not for
- Qwen3.8 27B at 4-bit or long agent contexts: that needs a 24 GB card
- Anyone who wants 70B models: none fit at any quant
- Buying the RTX 5080 at today's prices, unless you also game