Local AI hardware guide
Qwen3.8 27B VRAM Requirements: GPU, RAM, Quantization and Context Guide
Ollama lists Qwen3.8 27B Q4_K_M at 18GB and Q8_0 at 30GB. That makes a 16GB GPU too small for full weight residency; 24GB is a limited-headroom Q4 planning fit, while 32GB gives Q4 more room. These are package-size comparisons and AIPCFit estimates, not measured requirements or speed claims.
The short answer: the Ollama Q4_K_M package is listed at 18GB, so it is larger than a 16GB GPU's capacity. A 24GB card reaches AIPCFit's Q4 minimum-load planning estimate but sits just under its recommended estimate. A 32GB card clears that Q4 estimate. Q8_0 is listed at 30GB and remains a tight planning case even at 32GB. None of these comparisons guarantees runtime fit at every context length.
Start with the exact model record in the AIPCFit model catalog, compare a configuration in the VRAM calculator, or enter your hardware in the local PC checker.
What Qwen3.8 27B is
Qwen's model card identifies Qwen3.8-27B as a dense 27-billion-parameter vision-language model. It uses a 64-layer hybrid layout that combines Gated DeltaNet and gated attention. The model card lists 262,144 tokens as native context and says extension up to 1,000,000 tokens is possible with compatible long-context techniques. A model's advertised maximum is not a promise that a consumer GPU can hold that context in memory.
Q4_K_M and Q8_0 packages
The Q4_K_M and Q8_0 labels describe weight quantization formats in the Ollama GGUF variants. Q4_K_M uses a lower-bit K-quant recipe and has the smaller listed package; Q8_0 stores weights at 8-bit quantization and has a larger package. The 18GB and 30GB figures below are the sizes shown by the Ollama library, not a prediction of peak VRAM. The Ollama listings also show a 461M BF16 vision projector; its separate metadata entry should not be mistaken for a measured extra-memory requirement.
Qwen's Hugging Face repository contains the original Transformers-format weights and documents Transformers, vLLM, and SGLang serving. The size comparisons here refer specifically to Ollama's quantized GGUF tags; AIPCFit's catalog records those two variants for Ollama. A runtime's ability to load another community artifact should be checked for that exact model and file.
| Variant | Listed package size | Planning note |
|---|---|---|
| Q4_K_M | 18GB | Smaller of these two listed packages; still needs runtime and context headroom. |
| Q8_0 | 30GB | Larger package at 8-bit weight quantization; not a guarantee of better task results. |
How to read AIPCFit's memory estimates
AIPCFit uses each listed package size as a starting signal, then adds a planning margin: 15% for its minimum-load estimate and 35% for its recommended-memory estimate. For Q4_K_M, those calculations give 20.7GB and 24.3GB. For Q8_0, they give 34.5GB and 40.5GB. They are consistent catalog heuristics, not official Qwen requirements, runtime measurements, or context-specific VRAM calculations.
| GPU VRAM | Q4_K_M (18GB package) | Q8_0 (30GB package) | Ollama default context |
|---|---|---|---|
| 16GB | Does not fit entirely in dedicated VRAM; system-memory offload is a separate option. | Package is larger than VRAM; not a full-GPU fit. | 4K when Ollama detects below 24 GiB. |
| 24GB | AIPCFit fits planning status: above 20.7GB minimum, below 24.3GB recommended. | Package is larger than VRAM; not a full-GPU fit. | 4K below 24 GiB; 32K if it detects 24–48 GiB. |
| 32GB | AIPCFit good-fit planning status against the 24.3GB estimate. | AIPCFit tight planning status: above 30GB package size, below 34.5GB minimum estimate. | 32K if Ollama detects 24–48 GiB. |
These dedicated-VRAM comparisons do not rule out CPU or system-RAM offload when a package is larger than VRAM. Offloading is not the same as fitting all weights on the GPU, and its practicality depends on RAM, runtime support, and the rest of your workload. AIPCFit does not publish a Qwen3.8 minimum system-RAM requirement.
For examples from the GPU catalog, compare the 16GB RTX 5060 Ti overview, 24GB RTX 4090 overview, and 32GB RTX 5090 overview. The matching 16GB, 24GB, and 32GB VRAM guides show broader model choices for each tier.
Context length changes the fit
The 262K native context is a model limit, not a recommended setting for every PC. Ollama documents default contexts of 4K below 24 GiB VRAM and 32K from 24 to 48 GiB, and says increasing context increases memory use. Its guidance suggests at least 64K for coding tools and other long-context tasks; treat that as a workload setting to assess, not a Qwen3.8 hardware requirement. AIPCFit does not calculate Qwen3.8 memory by context length.
Begin with a context appropriate to your task, then check what the runtime actually allocated. Images and video use the model's vision path and can add workload-specific memory pressure; no exact extra amount is established by the catalog. If memory is tight, shorten context, close other GPU-heavy apps, or compare the smaller Q4_K_M package before considering offload.
Run and check the Ollama model
- Install Ollama for your operating system and check its current GPU and driver support.
- For the smaller listed variant, run `ollama run qwen3.8:27b-q4_K_M`. To use Q8_0 instead, run `ollama run qwen3.8:27b-q8_0`.
- Set or inspect the context length in Ollama settings. Do not assume the model card's full native context is allocated by default.
- Run `ollama ps` while the model is loaded and inspect the processor split and context columns. A CPU/GPU split indicates offload rather than full GPU residency.
- Use AIPCFit's PC checker to compare your GPU and system RAM with the intended variant, then test the prompts and context your workload needs.
AIPCFit currently has a generated compatibility page for the 24GB RTX 4090 Q4 planning path: Qwen3.8 27B on RTX 4090. It reports memory planning, not measured speed or a guarantee at a chosen context.
Want a result based on your own components? Check your PC with AIPCFit and compare the supported model configuration with your GPU, RAM, and runtime.
