RTX 4090 vs RTX 5090 for local LLMs
Assumes 32 GB of DDR5-5600 system RAM, Windows with this GPU driving the display, 8k context and an f16 KV cache.
What to know before choosing
Nemotron 3.5 Lightning 30B-A3B at Q4_K_M needs 25.1 GB, so in CanRun's count it fits fully only on the RTX 5090, where it runs great. The RTX 4090 keeps some experts in system RAM and runs it well, a cap from that split rather than slowness; both are theoretical. With an f16 KV cache, Qwen3.5 27B at 128k needs 26.4 GB: it runs great fully on the RTX 5090 but spills and runs slowly on the RTX 4090. A q8_0 KV cache fits it there with little margin — check ollama ps.
Each token reads the active weights from VRAM, so bandwidth sets the gap. In CanRun's estimate it is widest on dense models such as Gemma 4 31B; MoE models like Qwen3.6 35B-A3B gain less, because they carry a fixed per-token routing cost that bandwidth does not cut. Prompt reading follows FP16 compute, where the RTX 5090 leads by much less, so long prompts narrow the gap.
Llama 3.3 70B at Q4_K_M fits neither card: split with system RAM, it is too slow to use on the 24GB card and slow on the 32GB one. Only the 32GB card holds its heavy-loss IQ2_M file fully, where it runs great. The GeForce RTX 5090 Laptop has only 24 GB, so Nemotron keeps experts in system RAM there too, also a theoretical estimate.
Speed
The RTX 5090 generates text about 1.8× faster than the RTX 4090 (median of our estimates over the 46 models that fit entirely in memory on both at Q4_K_M or the closest available quant).
On Qwen3.5 35B-A3B the two estimate ranges overlap, so the real gap may be smaller.
- Reference model
- Qwen3.5 35B-A3B at Q4_K_M, the largest model that fits entirely in memory on both at Q4_K_M or the closest available quant
- RTX 4090
- est. 155.9 tok/s (124.7–187.1, calibrated estimate ±20%)
- RTX 5090
- est. 213.4 tok/s (170.7–256.0, calibrated estimate ±20%)
- Prompt processing on Qwen3.5 35B-A3B
- RTX 4090: about 9,387 tok/s (5,632–13,141, rough estimate ±40%); RTX 5090: about 11,910 tok/s (7,146–16,673, rough estimate ±40%)
Specs side by side
| Spec | RTX 4090 | RTX 5090 | Difference |
|---|---|---|---|
| Memory | 24 GB | 32 GB | +8 GB on the RTX 5090 |
| Memory bandwidth | 1,008.0 GB/s | 1,792.0 GB/s | 1.78× on the RTX 5090 |
| FP16 compute | 82.6 TFLOPS | 104.8 TFLOPS | 1.27× on the RTX 5090 |
| Launch year | 2022 | 2025 | — |
| Price | $2,755 street (as of 2026-08-09) | $4,200 street (as of 2026-08-09) | — |
| Public benchmark results | 13 results | 15 results | — |
Verdicts at a glance
Out of 58 models at Q4_K_M (or the closest available quant) with 8k context.
- RTX 4090Runs great46 modelsRuns well2 modelsRuns slowly0 modelsWon't run10 models
- RTX 5090Runs great48 modelsRuns well1 modelRuns slowly1 modelWon't run8 models
How to read the verdicts
- Runs great
- Fully on the GPU at 20 tok/s or more
- Runs well
- Fully on the GPU at 8–20 tok/s; MoE experts in system RAM at 20 tok/s or more; or at least 90% on the GPU at 8 tok/s or more
- Runs slowly
- 2–8 tok/s; CPU-only; less than 90% on the GPU; or MoE experts in system RAM below 20 tok/s
- Won't run
- Does not fit, or under 2 tok/s
Where the verdicts differ
4 models get a different verdict on the two.
Scroll sideways to see every column.
| Model | RTX 4090 verdict | RTX 4090 speed | RTX 5090 verdict | RTX 5090 speed |
|---|---|---|---|---|
| K2-Horizon MoVA 36B-A4BQ4_K_M | Runs well | est. 82.6 tok/s57.8–107.4theoretical estimate ±30% | Runs great | est. 149.6 tok/s104.7–194.5theoretical estimate ±30% |
| Nemotron 3.5 Lightning 30B-A3BQ4_K_M | Runs well | est. 126.5 tok/s88.5–164.4theoretical estimate ±30% | Runs great | est. 305.7 tok/s214.0–397.4theoretical estimate ±30% |
| Kolibri 1Q4_K_M | Won't runTry IQ4_XS: Runs well | est. 70.1 tok/s49.0–91.1theoretical estimate ±30% | Runs well | est. 83.3 tok/s58.3–108.3theoretical estimate ±30% |
| Llama 3.3 70BQ4_K_M | Won't runTry IQ4_XS: Runs slowly | est. 2.2 tok/s1.8–2.6calibrated estimate ±20% | Runs slowly | est. 2.7 tok/s2.2–3.3calibrated estimate ±20% |
Every model on both
Scroll sideways to see every column.
| Model | RTX 4090 verdict | RTX 4090 speed | RTX 5090 verdict | RTX 5090 speed |
|---|---|---|---|---|
| K2-Horizon MoVA 36B-A4BQ4_K_M | Runs well | est. 82.6 tok/s57.8–107.4theoretical estimate ±30% | Runs great | est. 149.6 tok/s104.7–194.5theoretical estimate ±30% |
| Qwen3.5 35B-A3BQ4_K_M | Runs great | est. 155.9 tok/s124.7–187.1calibrated estimate ±20% | Runs great | est. 213.4 tok/s170.7–256.0calibrated estimate ±20% |
| Qwen3.6 35B-A3BQ4_K_M | Runs great | est. 171.9 tok/s137.5–206.2calibrated estimate ±20% | Runs great | est. 229.8 tok/s183.9–275.8calibrated estimate ±20% |
| MiniCPM-V 4.7 35B-A3BQ4_K_M | Runs great | est. 181.6 tok/s145.3–217.9calibrated estimate ±20% | Runs great | est. 239.5 tok/s191.6–287.4calibrated estimate ±20% |
| EXAONE 4.5 33BQ4_K_M | Runs great | est. 34.3 tok/s30.2–38.4calibrated estimate ±12% | Runs great | est. 60.9 tok/s53.6–68.2calibrated estimate ±12% |
| DeepSeek R1 Distill Qwen 32BQ4_K_M | Runs great | est. 32.1 tok/s28.2–35.9calibrated estimate ±12% | Runs great | est. 57.0 tok/s50.2–63.9calibrated estimate ±12% |
| Qwen3 32BQ4_K_M | Runs great | 42.5 tok/s8k estimate 32.2 tok/s (28.3–36.1, calibrated estimate ±12%)measured (1 run, 512 context) | Runs great | 53.3 tok/s8k estimate 57.3 tok/s (50.4–64.1, calibrated estimate ±12%)measured (1 run, 512 context) |
| EXAONE 4.0 32BQ4_K_M | Runs great | est. 35.5 tok/s31.2–39.8calibrated estimate ±12% | Runs great | est. 63.1 tok/s55.5–70.7calibrated estimate ±12% |
| Nemotron 3.5 Lightning 30B-A3BQ4_K_M | Runs well | est. 126.5 tok/s88.5–164.4theoretical estimate ±30% | Runs great | est. 305.7 tok/s214.0–397.4theoretical estimate ±30% |
| Gemma 4 31BQ4_K_M | Runs great | est. 37.2 tok/s32.7–41.6calibrated estimate ±12% | Runs great | est. 66.1 tok/s58.1–74.0calibrated estimate ±12% |
| Qwen3 30B-A3B (2507)Q4_K_M | est. 146.7 tok/s117.4–176.1calibrated estimate ±20% | est. 195.0 tok/s156.0–234.1calibrated estimate ±20% | ||
| Muse Glimmer 30BQ4_K_M | Runs great | est. 40.5 tok/s35.6–45.4calibrated estimate ±12% | Runs great | est. 72.0 tok/s63.4–80.7calibrated estimate ±12% |
| Granite 4.2 30BQ4_K_M | Runs great | est. 35.5 tok/s31.3–39.8calibrated estimate ±12% | Runs great | est. 63.1 tok/s55.6–70.7calibrated estimate ±12% |
| Qwen3.5 27BQ4_K_M | Runs great | 44.8 tok/s8k estimate 40.0 tok/s (35.2–44.8, calibrated estimate ±12%)measured (2 runs, 512/40k context) | Runs great | est. 71.1 tok/s62.6–79.6calibrated estimate ±12% |
| Qwen3.6 27BQ4_K_M | Runs great | est. 40.7 tok/s35.8–45.5calibrated estimate ±12% | Runs great | est. 72.3 tok/s63.6–80.9calibrated estimate ±12% |
| Qwen3.8 27BQ4_K_M | Runs great | 49.8 tok/s8k estimate 39.3 tok/s (34.5–44.0, calibrated estimate ±12%)measured (1 run, 1k context) | Runs great | 81.8 tok/s8k estimate 69.8 tok/s (61.4–78.2, calibrated estimate ±12%)measured (1 run, 1k context) |
| Gemma 3 27BQ4_K_M | Runs great | est. 41.0 tok/s36.1–45.9calibrated estimate ±12% | Runs great | est. 72.8 tok/s64.1–81.6calibrated estimate ±12% |
| Gemma 4 26B-A4BQ4_K_M | Runs great | est. 142.6 tok/s114.1–171.1calibrated estimate ±20% | Runs great | 216.1 tok/s8k estimate 210.4 tok/s (168.3–252.4, calibrated estimate ±20%)measured (2 runs, 512/8k context) |
| Mistral Small 3.2 24BQ4_K_M | Runs great | est. 45.0 tok/s39.6–50.4calibrated estimate ±12% | Runs great | est. 80.0 tok/s70.4–89.6calibrated estimate ±12% |
| gpt-oss-20bMXFP4 | Runs greatgpt-oss-20b verdict on the RTX 4090 | 225.2 tok/s8k estimate 214.1 tok/s (171.3–256.9, calibrated estimate ±20%)measured (1 run, 2k context) | Runs greatgpt-oss-20b verdict on the RTX 5090 | 282.3 tok/s8k estimate 338.9 tok/s (271.1–406.7, calibrated estimate ±20%)measured (1 run, 2k context) |
| Kanana 1.5 15.7B-A3BQ4_K_M | Runs great | est. 164.4 tok/s131.5–197.3calibrated estimate ±20% | Runs great | est. 233.4 tok/s186.7–280.1calibrated estimate ±20% |
| Qwen3 14BQ4_K_M | Runs greatQwen3 14B verdict on the RTX 4090 | est. 68.2 tok/s60.0–76.4calibrated estimate ±12% | Runs greatQwen3 14B verdict on the RTX 5090 | est. 121.3 tok/s106.7–135.8calibrated estimate ±12% |
| DeepSeek R1 Distill Qwen 14BQ4_K_M | Runs great | est. 66.6 tok/s58.6–74.5calibrated estimate ±12% | Runs great | est. 118.3 tok/s104.1–132.5calibrated estimate ±12% |
| HyperCLOVA X SEED Think 14BQ4_K_M | Runs great | est. 69.2 tok/s48.4–90.0theoretical estimate ±30% | Runs great | est. 123.0 tok/s86.1–160.0theoretical estimate ±30% |
| Phi-4Q4_K_M | Runs great | est. 65.8 tok/s57.9–73.7calibrated estimate ±12% | Runs great | est. 116.9 tok/s102.9–131.0calibrated estimate ±12% |
| Gemma 3 12BQ4_K_M | Runs greatGemma 3 12B verdict on the RTX 4090 | est. 90.0 tok/s79.2–100.8calibrated estimate ±12% | Runs greatGemma 3 12B verdict on the RTX 5090 | est. 160.1 tok/s140.9–179.3calibrated estimate ±12% |
| Mellum2.1 12B-A2.5B ThinkingQ4_K_M | Runs great | est. 252.0 tok/s201.6–302.5calibrated estimate ±20% | Runs great | est. 334.9 tok/s267.9–401.8calibrated estimate ±20% |
| Gemma 4 12BQ4_K_M | Runs great | est. 97.3 tok/s85.6–108.9calibrated estimate ±12% | Runs great | est. 172.9 tok/s152.2–193.7calibrated estimate ±12% |
| Qwen3.5 9BQ4_K_M | Runs great | 127.5 tok/s8k estimate 114.9 tok/s (101.2–128.7, calibrated estimate ±12%)measured (2 runs, 512/40k context) | Runs great | est. 204.4 tok/s179.8–228.9calibrated estimate ±12% |
| K2-Horizon 7BQ4_K_M | Runs great | est. 103.8 tok/s91.3–116.3calibrated estimate ±12% | Runs great | est. 184.5 tok/s162.4–206.7calibrated estimate ±12% |
| Qwen3 8BQ4_K_M | Runs great | est. 113.1 tok/s99.5–126.7calibrated estimate ±12% | Runs great | 234.3 tok/s8k estimate 201.1 tok/s (177.0–225.2, calibrated estimate ±12%)measured (1 run, 512 context) |
| DeepSeek R1 0528 Qwen3 8BQ4_K_M | Runs great | est. 113.1 tok/s99.5–126.7calibrated estimate ±12% | Runs great | est. 201.1 tok/s177.0–225.2calibrated estimate ±12% |
| DeepSeek R1 Distill Llama 8BQ4_K_M | Runs great | est. 117.7 tok/s103.6–131.8calibrated estimate ±12% | Runs great | est. 209.3 tok/s184.2–234.4calibrated estimate ±12% |
| Kanana 1.5 8BQ4_K_M | Runs great | est. 117.7 tok/s103.6–131.8calibrated estimate ±12% | Runs great | est. 209.3 tok/s184.2–234.4calibrated estimate ±12% |
| Llama 3.1 8BQ4_K_M | Runs greatLlama 3.1 8B verdict on the RTX 4090 | 91.0 tok/s8k estimate 117.7 tok/s (103.6–131.8, calibrated estimate ±12%)measured (1 run, 4k context) | Runs greatLlama 3.1 8B verdict on the RTX 5090 | est. 209.3 tok/s184.2–234.4calibrated estimate ±12% |
| Gemma 4 E4BQ4_K_M | Runs great | est. 171.0 tok/s119.7–222.3theoretical estimate ±30% | Runs great | 234.0 tok/s8k estimate 257.7 tok/s (180.4–335.0, theoretical estimate ±30%)measured (1 run, 512 context) |
| Ling-3.0 TinyQ4_K_M | Runs great | est. 390.2 tok/s273.2–507.3theoretical estimate ±30% | Runs great | est. 485.1 tok/s339.6–630.7theoretical estimate ±30% |
| Gemma 4 E2BQ4_K_M | Runs great | est. 281.9 tok/s197.3–366.5theoretical estimate ±30% | Runs great | 341.8 tok/s8k estimate 386.7 tok/s (270.7–502.7, theoretical estimate ±30%)measured (1 run, 512 context) |
| Qwen3.5 4BQ4_K_M | Runs great | est. 227.7 tok/s200.4–255.1calibrated estimate ±12% | Runs great | est. 404.8 tok/s356.3–453.4calibrated estimate ±12% |
| Nanbeige4.2 3BQ4_K_M | Runs great | est. 115.2 tok/s80.6–149.7theoretical estimate ±30% | Runs great | est. 204.8 tok/s143.3–266.2theoretical estimate ±30% |
| Spark-X2.5 4BQ4_K_M | Runs great | est. 243.1 tok/s214.0–272.3calibrated estimate ±12% | Runs great | est. 432.3 tok/s380.4–484.1calibrated estimate ±12% |
| Llama 3.2 3BQ4_K_M | Runs great | est. 238.4 tok/s209.8–267.0calibrated estimate ±12% | Runs great | est. 423.9 tok/s373.0–474.7calibrated estimate ±12% |
| LFM2.5 2.6BQ4_K_M | Runs great | est. 391.1 tok/s273.8–508.4theoretical estimate ±30% | Runs great | est. 695.3 tok/s486.7–903.8theoretical estimate ±30% |
| MiniCPM5 2BQ4_K_M | Runs great | est. 369.0 tok/s324.7–413.3calibrated estimate ±12% | Runs great | est. 656.0 tok/s577.2–734.7calibrated estimate ±12% |
| HyperCLOVA X SEED 1.5BQ4_K_M | Runs great | est. 388.7 tok/s342.1–435.3calibrated estimate ±12% | Runs great | est. 691.0 tok/s608.1–773.9calibrated estimate ±12% |
| EXAONE 4.0 1.2BQ4_K_M | Runs great | est. 537.3 tok/s472.8–601.7calibrated estimate ±12% | Runs great | est. 955.1 tok/s840.5–1,069.8calibrated estimate ±12% |
| Llama 3.2 1BQ4_K_M | Runs great | est. 654.3 tok/s575.8–732.8calibrated estimate ±12% | Runs great | est. 1,163.2 tok/s1,023.6–1,302.7calibrated estimate ±12% |
| MiniCPM5 1BQ4_K_M | Runs great | est. 791.6 tok/s696.6–886.6calibrated estimate ±12% | Runs great | est. 1,407.3 tok/s1,238.5–1,576.2calibrated estimate ±12% |
| Kolibri 1Q4_K_M | Won't runTry IQ4_XS: Runs well | est. 70.1 tok/s49.0–91.1theoretical estimate ±30% | Runs well | est. 83.3 tok/s58.3–108.3theoretical estimate ±30% |
| Llama 3.3 70BQ4_K_M | Won't runTry IQ4_XS: Runs slowly | est. 2.2 tok/s1.8–2.6calibrated estimate ±20% | Runs slowly | est. 2.7 tok/s2.2–3.3calibrated estimate ±20% |
| GLM-5.3 FlashQ4_K_M | Won't run | — | Won't run | — |
| DeepSeek V4 Flash 0731MXFP4 | Won't run | — | Won't run | — |
| Solar Open 2 250BIQ4_XS | Won't run | — | Won't run | — |
| Ling-3.0 FlashQ4_K_M | Won't run | — | Won't run | — |
| Qwen3.5 122B-A10BQ4_K_M | Won't runTry UD-Q2_K_XL (heavy quality loss): Runs well | est. 42.2 tok/s29.5–54.8theoretical estimate ±30% | Won't runTry UD-Q2_K_XL (heavy quality loss): Runs well | est. 66.4 tok/s46.5–86.3theoretical estimate ±30% |
| Laguna S 2.1Q4_K_M | Won't run | — | Won't run | — |
| gpt-oss-120bMXFP4 | Won't rungpt-oss-120b verdict on the RTX 4090 | — | Won't rungpt-oss-120b verdict on the RTX 5090 | — |
| Solar Open 100BQ4_K_M | Won't runTry Q2_K (heavy quality loss): Runs slowly | est. 28.9 tok/s20.2–37.5theoretical estimate ±30% | Won't runTry IQ4_XS: Runs slowly | est. 21.9 tok/s15.3–28.4theoretical estimate ±30% |
Models with a thinking mode spend extra tokens thinking, so their speed thresholds are 1.5× stricter (30 / 12 / 3 tok/s). With thinking turned off, a model may do one grade better.
Public benchmarks
Measured results on the RTX 4090
| Model | Quant | Backend | Context | Prompt (tok/s) | Generation (tok/s) | Flags | Source | Measured |
|---|---|---|---|---|---|---|---|---|
| llama-2-7b | Q4_0 | llama.cpp | 512 | 11,993.0 | 186.2 | — | github.com | 2025-08-06 |
| gpt-oss-20b | MXFP4 | llama.cpp | 2k | 8,078.0 | 225.2 | — | github.com | 2025-08-20 |
| Qwen3 8B | Q4_K_XL | llama.cpp | 16k | — | 104.3 | — | hardware-corner.net | 2026-08-09 |
| Qwen3 30B-A3B (2507) | Q4_K_XL | llama.cpp | 16k | — | 139.7 | — | hardware-corner.net | 2026-08-09 |
| Llama 3.1 8B | Q4_K_M | llama.cpp | 4k | 6,697.0 | 91.0 | ctx-weighted | localscore.ai | 2026-08-09 |
| Qwen3.8 27B | Q4_K_M | llama.cpp | 1k | 3,073.0 | 49.8 | CUDA | github.com | 2026-09-12 |
| Qwen3.5 9B | Q4_K_M | ollama | 40k | — | 133.0 | — | github.com | 2026-10-12 |
| Qwen3.5 27B | Q4_K_M | ollama | 40k | — | 47.6 | — | github.com | 2026-10-12 |
| Qwen3.5 9B | Q4_K_M | llama.cpp | 512 | — | 122.0 | — | github.com | 2026-10-14 |
| Qwen3.5 27B | Q4_K_M | llama.cpp | 512 | — | 42.0 | — | github.com | 2026-10-14 |
| Qwen3 8B | Q4_K_XL | llama.cpp | 4k | 9,251.0 | 141.3 | CUDA | hardware-corner.net | 2026-03-01 |
| Qwen3 32B | Q4_K_XL | llama.cpp | 4k | 2,448.0 | 39.6 | CUDA | hardware-corner.net | 2026-03-01 |
| Qwen3 32B | Q4_K_M | llama.cpp | 512 | 2,302.0 | 42.5 | Vulkan | github.com | 2025-12-09 |
Measured results on the RTX 5090
| Model | Quant | Backend | Context | Prompt (tok/s) | Generation (tok/s) | Flags | Source | Measured |
|---|---|---|---|---|---|---|---|---|
| llama-2-7b | Q4_0 | llama.cpp | 512 | 14,073.0 | 290.0 | — | github.com | 2025-10-23 |
| gpt-oss-20b | MXFP4 | llama.cpp | 2k | 9,841.0 | 282.3 | — | github.com | 2025-08-20 |
| Qwen3 8B | Q4_K_XL | llama.cpp | 16k | — | 145.3 | — | hardware-corner.net | 2026-08-09 |
| Qwen3 30B-A3B (2507) | Q4_K_XL | llama.cpp | 16k | — | 141.6 | — | hardware-corner.net | 2026-08-09 |
| gpt-oss-120b | MXFP4 | llama.cpp | 4k | — | 9.6 | --n-cpu-moe 21, DDR4 server | hardware-corner.net | 2026-08-09 |
| Gemma 4 26B-A4B | Q4_K_M | llama.cpp | 512 | 13,588.0 | 223.8 | CUDA, KV q8_0 | github.com | 2026-05-24 |
| Gemma 4 26B-A4B | Q4_K_M | ollama | 8k | — | 208.4 | CUDA, FA off | github.com | 2026-04-05 |
| Qwen3.8 27B | Q4_K_M | llama.cpp | 1k | 3,972.0 | 81.8 | CUDA | github.com | 2026-09-12 |
| Qwen3 8B | Q4_K_XL | llama.cpp | 4k | 11,933.0 | 200.4 | CUDA | hardware-corner.net | 2026-03-02 |
| Qwen3 32B | Q4_K_XL | llama.cpp | 4k | 2,931.0 | 61.4 | CUDA | hardware-corner.net | 2025-12-09 |
| Qwen3 8B | Q4_K_M | llama.cpp | 512 | — | 234.3 | CUDA, KV q8_0 | github.com | 2026-10-07 |
| Qwen3 8B | Q8_0 | llama.cpp | 512 | 10,787.0 | 136.2 | CUDA | github.com | 2025-10-16 |
| Qwen3 32B | Q4_K_M | llama.cpp | 512 | 3,446.0 | 53.3 | CUDA | github.com | 2026-09-15 |
| Gemma 4 E4B | Q4_K_M | llama.cpp | 512 | — | 234.0 | Vulkan | github.com | 2026-09-08 |
| Gemma 4 E2B | Q4_K_M | llama.cpp | 512 | — | 341.8 | Vulkan | github.com | 2026-09-08 |
Public benchmarks we calibrate against. Their conditions (context, backend, flags) can differ from the estimates above.
Frequently asked questions
Which is faster for local LLMs, the RTX 4090 or the RTX 5090?
The RTX 5090 generates text about 1.8× faster than the RTX 4090 (median of our estimates over the 46 models that fit entirely in memory on both at Q4_K_M or the closest available quant). Estimated speed on Qwen3.5 35B-A3B at Q4_K_M: the RTX 4090 est. 155.9 tok/s (124.7–187.1, calibrated estimate ±20%); the RTX 5090 est. 213.4 tok/s (170.7–256.0, calibrated estimate ±20%).
What is the largest model each can run entirely in memory?
The RTX 4090 fits Qwen3.5 35B-A3B at Q4_K_M (22.63 GB file) entirely in 24 GB, at est. 155.9 tok/s (124.7–187.1, calibrated estimate ±20%). The RTX 5090 fits K2-Horizon MoVA 36B-A4B at Q4_K_M (22.37 GB file) entirely in 32 GB, at est. 149.6 tok/s (104.7–194.5, theoretical estimate ±30%).
How many local LLMs run well on each?
At Q4_K_M with 8k context, out of 58 tracked models: the RTX 4090 — 46 Runs great, 2 Runs well, 0 Runs slowly, 10 Won't run; the RTX 5090 — 48 Runs great, 1 Runs well, 1 Runs slowly, 8 Won't run. 4 models get a different verdict.
Speeds are estimates from memory bandwidth, calibrated against public benchmarks, and each one comes with an error band and a confidence label. Real results vary with drivers, backend, context length and thermals.