Qwen3.8-27B VRAM: 19.74 GiB

Qwen3.8-27B parameter count and which catalog GPUs fit

Qwen3.8-27B has 27,781,427,952 parameters in the catalog config.GGUF Q4_K_M (mradermacher) needs 16.43 GiB–19.74 GiB at 8,192 tokens, one sequence, FP16 KV, llama.cpp typical extras. The file is 15.66 GiB. The catalog note for this file says the optional mmproj is not included in that byte count, and file size is not peak VRAM.Architecture coverage is partial, so every tier below is a rough estimate.This table is the calculator against catalog memory budgets. It is not a benchmark, and software support stays unverified.

Discrete cards where it runs comfortably: A100 (80GB HBM2e), H100 (80GB), L40S (48GB GDDR6 with ECC), RTX 5090 (32 GB GDDR7). Discrete cards where it fits: RTX 4090 (24 GB GDDR6X), RX 7900 XTX (24 GB GDDR6), RTX 3090 (24 GB GDDR6X). Tight fit: RX 7900 XT (20 GB GDDR6). Does not fit: RTX 4060 Ti (16 GB GDDR6), RTX 4070 Ti SUPER (16 GB GDDR6X), RTX 4080 SUPER (16 GB GDDR6X), RTX 5080 (16 GB GDDR7), RX 7800 XT (16 GB GDDR6), RTX 5060 Ti (16 GB GDDR7), RTX 4080 (16 GB GDDR6X), RX 9070 (16 GB GDDR6), RTX 5070 Ti (16 GB GDDR7), RX 9070 XT (16 GB GDDR6), RTX 4070 (12 GB GDDR6X), RTX 3060 (12 GB GDDR6), RTX 5070 (12 GB GDDR7), RTX 3080 (12 GB GDDR6X), RTX 4070 SUPER (12 GB GDDR6X), RTX 3080 (10 GB GDDR6X), RTX 4060 (8 GB GDDR6), RTX 3060 (8 GB GDDR6), RTX 5060 Ti (8 GB GDDR7), RTX 3070 (8 GB GDDR6). Runs comfortably on unified memory (shared with the system, not discrete VRAM plus RAM): Apple M4 Max (128 GB unified memory), DGX Spark (GB10) (128 GB LPDDR5x coherent unified system memory), Ryzen AI Max+ 395 (128 GB LPDDR5x-8000 unified memory), Apple M2 Max (64 GB unified memory).

Allowlisted comparisons and recipes

• Can Qwen3.8-27B run on RTX 4090?
• Can Qwen3.8-27B run on DGX Spark (GB10)?
• Can Qwen3.8-27B run on Apple M2 Max?
• Can Qwen3.8-27B run on RTX 5080?
• Can Qwen3.8-27B run on RTX 5090?
• Qwen3.8-27B-NVFP4 on GeForce RTX 5090 32GB (vLLM MTP-3)

Allowlisted can-i-run pages pin the quant stored on that combo. This table uses the model's no-device auto-select, GGUF Q4_K_M (mradermacher). The recipe page is a sourced deployment note, not a measurement by this calculator.

How this estimate is calculated

Catalog config for Qwen/Qwen3.8-27B (revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0) is the Hugging Face config.json linked below. text_config records 64 layers, hidden size 5,120, 24 attention heads, 4 KV heads, head_dim 256, and max_position_embeddings 262,144. The hybrid spec is 16 groups of 3 Gated DeltaNet layers and 1 gated-attention layer (16 attention layers, 48 linear layers). KV uses those attention layers only. The all-layer GQA formula is not used. The no-device auto-select is GGUF Q4_K_M (mradermacher), file Qwen3.8-27B.Q4_K_M.gguf from mradermacher/Qwen3.8-27B-GGUF, 16,810,714,848 bytes. At 8,192 tokens, 1 sequence, FP16 KV, llama.cpp, the calculator estimates 16.43 GiB–19.74 GiB. This catalog has no calibration and no planned benchmark for this model. The range is a calculator estimate, not a measured run.

• config.json num_hidden_layers64
• config.json hidden_size5,120
• config.json num_attention_heads24
• config.json num_key_value_heads4
• config.json head_dim256
• config.json max_position_embeddings262,144
• catalog hybrid groups16
• Gated DeltaNet layers48
• gated-attention layers16
• linear_num_key_heads16
• linear_num_value_heads48
• linear_key_head_dim128
• linear_value_head_dim128
• linear_conv_kernel_dim4
• recurrent state dtypefp32
• total parameters27,781,427,952
• weight fileQwen3.8-27B.Q4_K_M.gguf (16,810,714,848 bytes)

Byte split for this setup

Weights, attention-layer KV, and the Gated DeltaNet state are separate. The state does not scale with context. File size is not the total.

• weight bytes16,810,714,848 bytes
• attention-layer KV bytes536,870,912 bytes
• Gated DeltaNet state bytes150,994,944 bytes
• conv state bytes7,864,320 bytes
• total16.43 GiB–19.74 GiB

Calculator notes for this setup

• Weight bytes come from the selected weight files only, not from every file in the repository.
• Counted files: Qwen3.8-27B.Q4_K_M.gguf.
• Download file size is not peak GPU memory. Runtime layout, allocator padding, and KV are extra.
• This checkpoint has a vision component. Text-only inference does not include vision encoder or projector bytes.
• Multi-token prediction / draft weights exist and are not included unless enabled.
• Hybrid layout: 16 groups of Gated DeltaNet plus gated attention (16 attention layers, 48 linear layers).
• Attention KV uses 16 gated-attention layers, 4 KV heads, head dim 256. Linear layers are not counted as extra KV layers.
• Gated DeltaNet recurrent state uses 1 sequences × 48 layers × 48 value heads × 128×128 × fp32 element size. It does not grow with context the way KV does, and is not treated as zero.
• Gated DeltaNet conv state uses 1 sequences × 48 layers × recorded channels × kernel 4 × fp32. Layout follows the published kernel dim, not a runtime dump.
• Using 4 KV heads, not 24 query heads.
• KV dtype is independent of weight quantization unless you change it.
• Recurrent and conv state dtype is independent of KV cache dtype. Changing KV precision does not silently change state precision.
• 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.
• Hybrid cache packing differs by engine (Transformers recurrent/conv vs vLLM mamba_cache_dtype / mamba_ssm_cache_dtype). This is a partial estimate, not a complete green result.
• llama.cpp compute buffers grow with context and batch. Values here are user-adjustable scenarios, not measured traces — unless a matching calibration is shown.
• No device is selected. The figures below are model memory demand. Add a device to judge whether it fits.

• config.json — revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0
• Qwen3.8-27B model card
• mradermacher/Qwen3.8-27B-GGUF weight repo
• Methodology
• KV cache calculator
• How much VRAM a local LLM needs

Runs on

• Can Qwen3.8-27B run on RTX 4090?
• Can Qwen3.8-27B run on DGX Spark (GB10)?
• Can Qwen3.8-27B run on Apple M2 Max?
• Can Qwen3.8-27B run on RTX 5080?
• Can Qwen3.8-27B run on RTX 5090?
• Qwen3.8-27B-NVFP4 on GeForce RTX 5090 32GB (vLLM MTP-3)

Memory layers in this estimate

Artifact bytes, raw KV, engine extras, and total demand are separate. This is not a runtime trace.

Weight artifact

15.66 GiB (16,810,714,848 bytes)

Raw KV cache

0.50 GiB (536,870,912 bytes)

Engine extras / workspace

0.27 GiB–3.58 GiB

Total estimated demand

16.43 GiB–19.74 GiB

Published artifacts

Memory by quantization and context

Concurrent sequences = 1. KV defaults to FP16. The range includes documented runtime extras for the listed engine. These figures use the same core as the interactive calculator.

Raw KV contribution

Related pages

• KV cache calculator
• 4-bit, 8-bit, and FP16 memory
• File size is not runtime VRAM
• Methodology
• Can Qwen3.8-27B run on RTX 4090?
• Can Qwen3.8-27B run on DGX Spark?
• Can Qwen3.8-27B run on Apple M2 Max 64GB unified (MacBook Pro / Mac Studio)?
• All can-i-run checks

Sources

• Qwen3.8-27B model card
• config.json — revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0
• safetensors index — metadata.total_size=55562855904
• Artifact repo unsloth/Qwen3.8-27B-GGUF
• Artifact repo mradermacher/Qwen3.8-27B-GGUF

Assumptions that affect this result

• Requested context is within the recorded native window of 262144 tokens.
• Full-resident weight estimates use total parameters, including inactive MoE experts.
• 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: Qwen3.8-27B.Q4_K_M.gguf.
• This checkpoint has a vision component. Text-only inference does not include vision encoder or projector bytes.
• Multi-token prediction / draft weights exist and are not included unless enabled.
• Context budget is the number of cached tokens per sequence (prompt plus reserved generation).

Calculator