Can an RTX 4090 run Qwen3.8-27B?
Yes: an RTX 4090 runs Qwen3.8 27B at 44.2 tok/s at Q4_K_M, measured on llama.cpp with 8K context in October 2026.
8 signed runs (7 default settings, 1 MTP speculative decoding on) on llm-speed. Single-request decode speed in tokens per second; the short-chat workload unless noted. 6 of 8 runs were measured by llm-speed (@llmspeed); the other 2 are community-submitted (signed on the submitter's own machine, not measured by llm-speed, and not used for llm-speed verdicts). Runs measured by llm-speed are credited to @llmspeed. Verdicts and headline numbers use only llm-speed's own runs and fit tests.
Measured tok/s by quant and runtime
| Quant | Runtime | Build | Decode tok/s (best run) | Workload | Runs | Source |
|---|---|---|---|---|---|---|
| Q4_K_M | llama.cpp | commit a37cc4610 | 44.2 | chat-short | 6 | Measured by llm-speed@llmspeed |
| Q4_K_M | llama.cpp | commit df18bee | 43.1 | chat-short | 1 | Community-submitted |
| Q4_K_MMTP speculative decoding on (counted separately) | llama.cpp | commit df18bee | 84.5 | chat-short | 1 | Community-submitted |
Fastest quant that fits: Q4_K_M on llama.cpp, 44.2 tok/s (measured by llm-speed; it fits because it ran on this hardware). Quants without a run are not listed.
VRAM used: 18.7 GiB at Q4_K_M (r_tt9io-8lu9p) peak, 8K context, of 24.0 GiB (nvidia-smi peak).
Every measured row
| Run | Date (UTC) | Workload | Quant | Runtime / build | Prompt tokens | Output tokens | Decode tok/s | Prefill tok/s | TTFT ms | VRAM GiB | Source |
|---|---|---|---|---|---|---|---|---|---|---|---|
| r_7j8rpk9mwhg | 2026-10-14 | chat-shortMTP speculative decoding on | Q4_K_M | llama.cpp · commit df18bee | 154 | 256 | 84.5 | 299.3 | 515 | – | Community-submitted |
| r_tt9io-8lu9p | 2026-10-22 | chat-short | Q4_K_M | llama.cpp · commit a37cc4610 | 154 | 256 | 44.2 | 691.5 | 223 | 18.7 peak | Measured by llm-speed@llmspeed |
| r_k1hhpa9io2o | 2026-10-22 | chat-short | Q4_K_M | llama.cpp · commit a37cc4610 | 154 | 256 | 43.4 | 442.4 | 348 | – | Measured by llm-speed@llmspeed |
| r_7cap38a021- | 2026-10-22 | chat-short | Q4_K_M | llama.cpp · commit a37cc4610 | 154 | 256 | 43.4 | 448.3 | 344 | – | Measured by llm-speed@llmspeed |
| r_cthpfm82n1r | 2026-10-22 | chat-short | Q4_K_M | llama.cpp · commit a37cc4610 | 154 | 256 | 43.4 | 472.6 | 326 | – | Measured by llm-speed@llmspeed |
| r_fttw19bkyio | 2026-10-22 | chat-short | Q4_K_M | llama.cpp · commit a37cc4610 | 154 | 256 | 43.3 | 461.3 | 334 | 18.7 peak | Measured by llm-speed@llmspeed |
| r_11kn1epxewh | 2026-10-14 | chat-short | Q4_K_M | llama.cpp · commit df18bee | 154 | 256 | 43.1 | 428.6 | 359 | – | Community-submitted |
| r_--ul19zazqg | 2026-10-22 | chat-short | Q4_K_M | llama.cpp · commit a37cc4610 | 154 | 256 | 43.1 | 514.7 | 299 | – | Measured by llm-speed@llmspeed |
| r_7j8rpk9mwhg | 2026-10-14 | chat-longMTP speculative decoding on | Q4_K_M | llama.cpp · commit df18bee | 3224 | 1024 | 73.1 | 1585.8 | 2033 | – | Community-submitted |
| r_tt9io-8lu9p | 2026-10-22 | chat-long | Q4_K_M | llama.cpp · commit a37cc4610 | 3224 | 1024 | 43.7 | 2412.3 | 1336 | 18.7 peak | Measured by llm-speed@llmspeed |
| r_--ul19zazqg | 2026-10-22 | chat-long | Q4_K_M | llama.cpp · commit a37cc4610 | 3224 | 1024 | 43.0 | 2015.4 | 1600 | – | Measured by llm-speed@llmspeed |
| r_fttw19bkyio | 2026-10-22 | chat-long | Q4_K_M | llama.cpp · commit a37cc4610 | 3224 | 1024 | 43.0 | 1992.2 | 1618 | 18.7 peak | Measured by llm-speed@llmspeed |
| r_k1hhpa9io2o | 2026-10-22 | chat-long | Q4_K_M | llama.cpp · commit a37cc4610 | 3224 | 1024 | 43.0 | 2008.7 | 1605 | – | Measured by llm-speed@llmspeed |
| r_cthpfm82n1r | 2026-10-22 | chat-long | Q4_K_M | llama.cpp · commit a37cc4610 | 3224 | 1024 | 42.9 | 1958.6 | 1646 | – | Measured by llm-speed@llmspeed |
| r_7cap38a021- | 2026-10-22 | chat-long | Q4_K_M | llama.cpp · commit a37cc4610 | 3224 | 872 | 42.9 | 1979.3 | 1629 | – | Measured by llm-speed@llmspeed |
| r_11kn1epxewh | 2026-10-14 | chat-long | Q4_K_M | llama.cpp · commit df18bee | 3224 | 1024 | 42.6 | 1915.1 | 1684 | – | Community-submitted |
| r_7j8rpk9mwhg | 2026-10-14 | agent-traceMTP speculative decoding on | Q4_K_M | llama.cpp · commit df18bee | 3496 | 840 | 77.6 | 2912.1 | 787 | – | Community-submitted |
| r_tt9io-8lu9p | 2026-10-22 | agent-trace | Q4_K_M | llama.cpp · commit a37cc4610 | 3496 | 840 | 43.4 | 4710.8 | 458 | 18.7 peak | Measured by llm-speed@llmspeed |
| r_cthpfm82n1r | 2026-10-22 | agent-trace | Q4_K_M | llama.cpp · commit a37cc4610 | 3496 | 840 | 42.6 | 3858.7 | 583 | – | Measured by llm-speed@llmspeed |
| r_--ul19zazqg | 2026-10-22 | agent-trace | Q4_K_M | llama.cpp · commit a37cc4610 | 3496 | 840 | 42.5 | 4207.8 | 545 | – | Measured by llm-speed@llmspeed |
| r_fttw19bkyio | 2026-10-22 | agent-trace | Q4_K_M | llama.cpp · commit a37cc4610 | 3496 | 840 | 42.4 | 3792.7 | 592 | 18.7 peak | Measured by llm-speed@llmspeed |
| r_k1hhpa9io2o | 2026-10-22 | agent-trace | Q4_K_M | llama.cpp · commit a37cc4610 | 3496 | 840 | 42.2 | 3873.7 | 576 | – | Measured by llm-speed@llmspeed |
| r_7cap38a021- | 2026-10-22 | agent-trace | Q4_K_M | llama.cpp · commit a37cc4610 | 3496 | 840 | 42.1 | 3815.6 | 581 | – | Measured by llm-speed@llmspeed |
| r_11kn1epxewh | 2026-10-14 | agent-trace | Q4_K_M | llama.cpp · commit df18bee | 3496 | 840 | 42.1 | 3644.0 | 629 | – | Community-submitted |
| r_7j8rpk9mwhg | 2026-10-14 | concurrent-decodeMTP speculative decoding on | Q4_K_M | llama.cpp · commit df18bee | 1024 | 256 | 87.0 | – | – | – | Community-submitted |
| r_tt9io-8lu9p | 2026-10-22 | concurrent-decode | Q4_K_M | llama.cpp · commit a37cc4610 | 1024 | 256 | 44.1 | – | – | 18.7 peak | Measured by llm-speed@llmspeed |
| r_fttw19bkyio | 2026-10-22 | concurrent-decode | Q4_K_M | llama.cpp · commit a37cc4610 | 1024 | 256 | 43.4 | – | – | 18.7 peak | Measured by llm-speed@llmspeed |
| r_k1hhpa9io2o | 2026-10-22 | concurrent-decode | Q4_K_M | llama.cpp · commit a37cc4610 | 1024 | 256 | 43.3 | – | – | – | Measured by llm-speed@llmspeed |
| r_7cap38a021- | 2026-10-22 | concurrent-decode | Q4_K_M | llama.cpp · commit a37cc4610 | 1024 | 256 | 43.3 | – | – | – | Measured by llm-speed@llmspeed |
| r_cthpfm82n1r | 2026-10-22 | concurrent-decode | Q4_K_M | llama.cpp · commit a37cc4610 | 1024 | 256 | 43.3 | – | – | – | Measured by llm-speed@llmspeed |
| r_--ul19zazqg | 2026-10-22 | concurrent-decode | Q4_K_M | llama.cpp · commit a37cc4610 | 1024 | 256 | 43.0 | – | – | – | Measured by llm-speed@llmspeed |
| r_11kn1epxewh | 2026-10-14 | concurrent-decode | Q4_K_M | llama.cpp · commit df18bee | 1024 | 256 | 43.0 | – | – | – | Community-submitted |
Questions
Can an RTX 4090 run Qwen3.8-27B?
Yes: an RTX 4090 runs Qwen3.8 27B at 44.2 tok/s at Q4_K_M, measured on llama.cpp with 8K context in October 2026.
How many tokens per second does Qwen3.8-27B get on an RTX 4090?
Decode speed measured by llm-speed, by quantization and runtime: Q4_K_M on llama.cpp: 44.2 tok/s. Based on 6 runs measured by llm-speed; single-request chat-short unless noted. 2 community-submitted runs are listed on the page, labelled, and not counted here.
What is the fastest quant of Qwen3.8-27B that fits on an RTX 4090?
Q4_K_M on llama.cpp was the fastest measured setting at 44.2 tok/s; it fit because it ran. Other quants are listed only where a run measured them.
How much VRAM does Qwen3.8-27B use on an RTX 4090?
nvidia-smi recorded a peak of 18.7 GiB of 24.0 GiB at Q4_K_M with 8K context (run r_tt9io-8lu9p).
Sources and related pages
- Source runs: r_7j8rpk9mwhg, r_tt9io-8lu9p, r_k1hhpa9io2o, r_7cap38a021-, r_cthpfm82n1r, r_fttw19bkyio, r_11kn1epxewh, r_--ul19zazqg
- Qwen3.8-27B: every measured GPU and Mac
- RTX 4090: every measured model
- VRAM calculator for Qwen3.8-27B on RTX 4090
- RTX 5090 vs RTX 4090
- RTX 4090 vs RTX 3090
- RTX 4090 vs M4 Max
- How runs are measured · open dataset (CC BY 4.0)