{ "checkedAt": "2026-11-02", "architectureSource": "https://huggingface.co/Qwen/Qwen3.8-27B/blob/main/config.json", "config": { "architectures": [ "Qwen3_5ForConditionalGeneration" ], "image_token_id": 248056, "language_model_only": false, "model_type": "qwen3_5", "text_config": { "attention_bias": false, "attention_dropout": 0.0, "attn_output_gate": true, "bos_token_id": 248044, "dtype": "bfloat16", "eos_token_id": 248044, "full_attention_interval": 4, "head_dim": 256, "hidden_act": "silu", "hidden_size": 5120, "initializer_range": 0.02, "intermediate_size": 17408, "layer_types": [ "linear_attention", "linear_attention", "linear_attention", "full_attention", "linear_attention", "linear_attention", "linear_attention", "full_attention", "linear_attention", "linear_attention", "linear_attention", "full_attention", "linear_attention", "linear_attention", "linear_attention", "full_attention", "linear_attention", "linear_attention", "linear_attention", "full_attention", "linear_attention", "linear_attention", "linear_attention", "full_attention", "linear_attention", "linear_attention", "linear_attention", "full_attention", "linear_attention", "linear_attention", "linear_attention", "full_attention", "linear_attention", "linear_attention", "linear_attention", "full_attention", "linear_attention", "linear_attention", "linear_attention", "full_attention", "linear_attention", "linear_attention", "linear_attention", "full_attention", "linear_attention", "linear_attention", "linear_attention", "full_attention", "linear_attention", "linear_attention", "linear_attention", "full_attention", "linear_attention", "linear_attention", "linear_attention", "full_attention", "linear_attention", "linear_attention", "linear_attention", "full_attention", "linear_attention", "linear_attention", "linear_attention", "full_attention" ], "linear_conv_kernel_dim": 4, "linear_key_head_dim": 128, "linear_num_key_heads": 16, "linear_num_value_heads": 48, "linear_value_head_dim": 128, "mamba_ssm_dtype": "float32", "max_position_embeddings": 262144, "model_type": "qwen3_5_text", "mtp_num_hidden_layers": 1, "mtp_use_dedicated_embeddings": false, "num_attention_heads": 24, "num_hidden_layers": 64, "num_key_value_heads": 4, "output_gate_type": "swish", "pad_token_id": null, "partial_rotary_factor": 0.25, "rms_norm_eps": 1e-06, "rope_parameters": { "mrope_interleaved": true, "mrope_section": [ 11, 11, 10 ], "partial_rotary_factor": 0.25, "rope_theta": 10000000, "rope_type": "default" }, "tie_word_embeddings": false, "use_cache": true, "vocab_size": 248320 }, "tie_word_embeddings": false, "transformers_version": "5.8.0.dev0", "video_token_id": 248057, "vision_config": { "deepstack_visual_indexes": [], "depth": 27, "hidden_act": "gelu_pytorch_tanh", "hidden_size": 1152, "in_channels": 3, "initializer_range": 0.02, "intermediate_size": 4304, "model_type": "qwen3_5", "num_heads": 16, "num_position_embeddings": 2304, "out_hidden_size": 5120, "patch_size": 16, "spatial_merge_size": 2, "temporal_patch_size": 2 }, "vision_end_token_id": 248054, "vision_start_token_id": 248053 }, "model": { "id": "qwen3827", "name": "Qwen 3.8 · 27B", "p": 27.32, "layers": 16, "kv": 4, "dim": 256, "maxContext": 262144, "hybrid": true, "stateGB": 0.158859264, "weightGB": { "4": 16.464440224, "6": 21.983677344, "8": 29.047086048 }, "source": "https://huggingface.co/Qwen/Qwen3.8-27B" }, "weightSource": "https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/tree/main", "files": [ { "path": "Qwen3.8-27B-UD-Q4_K_M.gguf", "bytes": 16464440224 }, { "path": "Qwen3.8-27B-UD-Q6_K.gguf", "bytes": 21983677344 }, { "path": "Qwen3.8-27B-Q8_0.gguf", "bytes": 29047086048 }, { "path": "mmproj-BF16.gguf", "bytes": 931146432 } ], "memoryMeasurements": [ { "context": 4096, "measuredGB": 18 }, { "context": 8192, "measuredGB": 18 }, { "context": 16384, "measuredGB": 19 }, { "context": 32768, "measuredGB": 20 }, { "context": 65536, "measuredGB": 22 }, { "context": 131072, "measuredGB": 26 }, { "context": 262144, "measuredGB": 34 } ], "memoryMeasurementSource": "https://www.hardware-corner.net/qwen3-8-27b-hardware-tests/", "memoryMeasurementNote": "Q4_K_S, andere Datei als das Laborprofil. Gerundete Laufzeitbelegung; kein MLX- oder universelles Kontextlimit.", "sources": { "qwen38hc": { "name": "Hardware Corner · Qwen3.8 · eigene Messungen", "url": "https://www.hardware-corner.net/qwen3-8-27b-hardware-tests/", "date": "17.08.2026 · Abruf 02.11.2026", "setup": "RTX 5090 32 GB · Ubuntu 24.04 · CUDA 12.8 · Treiber 595.58.03 · llama.cpp 153d324bc (10364).", "protocol": "Q4_K_S, 16,68 GiB laut Quelle · Flash Attention an · vollständiger GPU-Offload · MTP aus. Kontextstufen wie im Artikel; genaue PP-/TG-Längen nicht genannt. VRAM-Messwerte gerundet; Host-CPU/RAM und Mac-Bauform nicht belegt. Deshalb keine Übertragung der dortigen M5-Max-Werte auf einen Desktop." }, "qwen38macstories": { "name": "Federico Viticci · MacStories · Desktopvergleich", "url": "https://www.macstories.net/stories/m5-ultra-mac-studio-review-the-dream-mac-for-local-ai-agents/#m5-ultra-vs-rtx-5090", "date": "Abruf 02.11.2026", "setup": "Mac Studio M5 Ultra 80 GPU / 256 GB und M3 Ultra / 512 GB: oMLX 0.7.0.dev2. RTX 5090 32 GB, Ryzen 9 9950X3D, 96 GB RAM, Windows 11 Pro: LM Studio 0.4.24, CUDA-12-Runtime 2.41.0, Treiber 616.92.", "protocol": "Identische Aufgabe, einzelne Läufe: 6.091 bzw. 63.445 Prompt-Tokens; Output-Limit 2.048, Thinking und MTP aus. Apple MLX oQ4e, RTX GGUF Q4_K_M: verschiedene Quantisierungen. 64K: M5 Ultra hat 2.048 gecachte Tokens; RTX-PP nicht berichtet. Ganze Systeme, kein isolierter Chipvergleich; kein Messnachweis für unseren 32-GB-RAM-PC." } }, "rows": [ { "family": "rtx", "hardware": "RTX 5090 · 32 GB", "host": "Ubuntu 24.04 · Host-CPU/RAM nicht angegeben", "model": "Qwen 3.8 27B", "quant": "Q4_K_S · 16,68 GiB laut Quelle", "context": 4096, "contextLabel": "4K Kontext laut Quelle", "testNote": "FA an · MTP aus · PP-/TG-Länge nicht genannt", "tg": 74.83, "pp": 3749.51, "source": "qwen38hc", "backend": "llama.cpp · CUDA + FA", "build": "153d324bc (10364)" }, { "family": "rtx", "hardware": "RTX 5090 · 32 GB", "host": "Ubuntu 24.04 · Host-CPU/RAM nicht angegeben", "model": "Qwen 3.8 27B", "quant": "Q4_K_S · 16,68 GiB laut Quelle", "context": 8192, "contextLabel": "8K Kontext laut Quelle", "testNote": "FA an · MTP aus · PP-/TG-Länge nicht genannt", "tg": 52.23, "pp": 3592.89, "source": "qwen38hc", "backend": "llama.cpp · CUDA + FA", "build": "153d324bc (10364)" }, { "family": "rtx", "hardware": "RTX 5090 · 32 GB", "host": "Ubuntu 24.04 · Host-CPU/RAM nicht angegeben", "model": "Qwen 3.8 27B", "quant": "Q4_K_S · 16,68 GiB laut Quelle", "context": 65536, "contextLabel": "64K Kontext laut Quelle", "testNote": "FA an · MTP aus · PP-/TG-Länge nicht genannt", "tg": 26.22, "pp": 717.76, "source": "qwen38hc", "backend": "llama.cpp · CUDA + FA", "build": "153d324bc (10364)" }, { "family": "rtx", "hardware": "RTX 5090 · 32 GB", "host": "Ryzen 9 9950X3D · 96 GB RAM · Windows 11 Pro", "model": "Qwen 3.8 27B", "quant": "GGUF Q4_K_M", "context": null, "contextLabel": "6.091 Prompt / max. 2.048 Output", "testNote": "Ein Lauf · Thinking / MTP aus · verschiedene Quants", "tg": 59, "pp": 3031, "source": "qwen38macstories", "backend": "LM Studio · CUDA", "build": "LM Studio 0.4.24 / Runtime 2.41.0" }, { "family": "apple", "hardware": "Mac Studio M5 Ultra · 80 GPU · 256 GB", "host": "macOS 27.0", "model": "Qwen 3.8 27B", "quant": "MLX oQ4e", "context": null, "contextLabel": "6.091 Prompt / max. 2.048 Output", "testNote": "Ein Lauf · Thinking / MTP aus · verschiedene Quants", "tg": 48, "pp": 1701, "source": "qwen38macstories", "backend": "oMLX · MLX", "build": "oMLX 0.7.0.dev2" }, { "family": "apple", "hardware": "Mac Studio M3 Ultra · 512 GB · Referenz", "host": "macOS 27.0 · ältere Desktop-Referenz", "model": "Qwen 3.8 27B", "quant": "MLX oQ4e", "context": null, "contextLabel": "6.091 Prompt / max. 2.048 Output", "testNote": "Ein Lauf · Thinking / MTP aus · verschiedene Quants", "tg": 31, "pp": 414, "source": "qwen38macstories", "backend": "oMLX · MLX", "build": "oMLX 0.7.0.dev2" }, { "family": "rtx", "hardware": "RTX 5090 · 32 GB", "host": "Ryzen 9 9950X3D · 96 GB RAM · Windows 11 Pro", "model": "Qwen 3.8 27B", "quant": "GGUF Q4_K_M", "context": null, "contextLabel": "63.445 Prompt / max. 2.048 Output", "testNote": "Ein Lauf · Thinking / MTP aus · verschiedene Quants", "tg": 51, "pp": null, "source": "qwen38macstories", "backend": "LM Studio · CUDA", "build": "LM Studio 0.4.24 / Runtime 2.41.0" }, { "family": "apple", "hardware": "Mac Studio M5 Ultra · 80 GPU · 256 GB", "host": "macOS 27.0 · 2.048 Tokens Cache-Treffer", "model": "Qwen 3.8 27B", "quant": "MLX oQ4e", "context": null, "contextLabel": "63.445 Prompt / max. 2.048 Output", "testNote": "Ein Lauf · Thinking / MTP aus · verschiedene Quants", "tg": 39, "pp": 1501, "source": "qwen38macstories", "backend": "oMLX · MLX", "build": "oMLX 0.7.0.dev2" }, { "family": "apple", "hardware": "Mac Studio M3 Ultra · 512 GB · Referenz", "host": "macOS 27.0 · ältere Desktop-Referenz", "model": "Qwen 3.8 27B", "quant": "MLX oQ4e", "context": null, "contextLabel": "63.445 Prompt / max. 2.048 Output", "testNote": "Ein Lauf · Thinking / MTP aus · verschiedene Quants", "tg": 24, "pp": 339, "source": "qwen38macstories", "backend": "oMLX · MLX", "build": "oMLX 0.7.0.dev2" } ] }