Mixtral-8x7B-Instruct-v0.1 on RTX 5090: Tight fit (Sep 2026)
Conditional memory judgement for one configuration. Change context, quant, or engine and the conclusion can change. Configuration used on this page: GGUF Q4_K_M (mradermacher), 8,192 cached tokens, 1 sequence, engine llama.cpp. Status: Tight. Software: Not verified..
Verdict on GeForce RTX 5090 32GB (desktop)
Only just: Q4_K_M is the one file that gets in. It needs 27.6 GiB–28.6 GiB at 8,192 tokens against 30.0 GiB of usable VRAM (after the driver reserve). Close other GPU apps first, and expect to cut context if the upper estimate is hit.
With Q4_K_M, the longest context that still fits at all (tight, no headroom) is about 27,691 tokens of a 32,768-token native window, one sequence, fp16 KV cache.
| GGUF quant | File weights | KV cache | Estimated total | Verdict |
|---|---|---|---|---|
| Q8_0 | 46.2 GiB | 1.0 GiB | 47.3 GiB–48.7 GiB | Does not fit |
| Q6_K | 35.7 GiB | 1.0 GiB | 36.9 GiB–38.0 GiB | Does not fit |
| Q5_K_M | 30.9 GiB | 1.0 GiB | 32.1 GiB–33.1 GiB | Does not fit |
| Q4_K_M (recommended) | 26.5 GiB | 1.0 GiB | 27.6 GiB–28.6 GiB | Tight |
Engine llama.cpp, one sequence, fp16 KV. File weights are the published GGUF sizes in the catalog. Totals come from the same calculator that powers the tool below; they are estimates, not a measurement of this setup.
Fixed page configuration
GGUF Q4_K_M needs 27.6 GiB-28.6 GiB at 8,192 tokens against 30.00 GiB net on GeForce RTX 5090 32GB (desktop), a tight fit. Q8_0, Q6_K, Q5_K_M are tight or over budget.
- ModelMixtral-8x7B-Instruct-v0.1
- HardwareGeForce RTX 5090 32GB (desktop)
- QuantizationGGUF Q4_K_M (mradermacher)
- Context length (cached tokens)8,192
- Engine
llama.cpp - Concurrent sequences1
- KV cache typefp16
Technical identifiers and share data
- Model
mixtral-8x7b-instruct-v0.1 - Hardware
nvidia-geforce-rtx-5090-32gb-desktop - Quantization
gguf-q4_k_m - Share fragment for this identity
v2.eyJzY2hlbWFWZXJzaW9uIjoyLCJjYWxjdWxhdG9yVmVyc2lvbiI6IjAuMy4wIiwiZGF0YXNldFZlcnNpb24iOiIyMDI2LTEwLTAyLjEiLCJjb25maWciOnsibW9kZWxJZCI6Im1peHRyYWwtOHg3Yi1pbnN0cnVjdC12MC4xIiwiaGFyZHdhcmVJZCI6Im52aWRpYS1nZWZvcmNlLXJ0eC01MDkwLTMyZ2ItZGVza3RvcCIsImRldmljZUNvdW50IjoxLCJxdWFudGl6YXRpb25JZCI6ImdndWYtcTRfa19tIiwiY29udGV4dFRva2VucyI6ODE5MiwiY29uY3VycmVudFNlcXVlbmNlcyI6MSwiZW5naW5lIjoibGxhbWEuY3BwIiwia3ZUeXBlIjoiZnAxNiJ9fQ
Memory breakdown
- Weight artifact
- 26.49 GiB (28,448,468,384 bytes)
- Raw KV cache
- 1.00 GiB (1,073,741,824 bytes)
- Engine extras / workspace
- 0.13 GiB–1.10 GiB
- Total estimated demand
- 27.62 GiB–28.59 GiB
- Net calculator budget
- 30.00 GiB
- OS / display reserve
- 2.00 GiB
- Remaining vs net budget
- 1.41 GiB–2.38 GiB
28.59 GiB high estimate vs 30.00 GiB net
Enough memory does not mean this engine, quant, and OS will run. Support does not mean it will be fast.
Other published files on this device
Same model, device and context (8,192 tokens, 1 sequence, FP16 KV). Each row is one published weight-file set with its usual engine, sorted by memory.
| Quantization | Engine | Estimated memory | Fit |
|---|---|---|---|
| BF16 safetensors | vllm | 88.62 GiB–90.52 GiB | Does not fit |
| GGUF Q8_0 (mradermacher) | llama.cpp | 47.34 GiB–48.71 GiB | Does not fit |
| GGUF Q6_K (mradermacher) | llama.cpp | 36.87 GiB–38.03 GiB | Does not fit |
| GGUF Q5_K_M (mradermacher) | llama.cpp | 32.07 GiB–33.14 GiB | Does not fit |
| GGUF Q4_K_M (mradermacher) This page | llama.cpp | 27.62 GiB–28.59 GiB | Tight fit |
Assumptions that affect this result
- Requested context is within the recorded native window of 32768 tokens.
- Full-resident weight estimates use total parameters, including inactive MoE experts.
- This MoE lists 46702792704 total parameters and 12900000000 active parameters. Active-parameter compute is not used as the weight-residency figure.
- Weight bytes come from the selected weight files only, not from every file in the repository.
- Download file size is not peak GPU memory. Runtime layout, allocator padding, and KV are extra.
- Counted files: Mixtral-8x7B-Instruct-v0.1.Q4_K_M.gguf.
- Context budget is the number of cached tokens per sequence (prompt plus reserved generation).
- Concurrent sequences share one model replica; weights are not multiplied by sequence count.
Related pages
Sources
- Mixtral-8x7B-Instruct-v0.1 model card
- config.json — revision eba92302a2861cdc0098cc54bc9f17cb2c47eb61
- safetensors index — metadata.total_size=93405585408
- Artifact repo mradermacher/Mixtral-8x7B-Instruct-v0.1-GGUF
- NVIDIA GeForce RTX 5090 product page — Standard Memory Config 32 GB GDDR7
Calculator
This combo page keeps the route model and device. Customize to change them on the main calculator; the share then keeps that identity.
Loading catalog snapshot…
Search or pick a catalog checkpoint to see memory demand.