Fleet and fit matrix
Which model runs on which machine. The Bosgame M5 columns come first because that is the box being planned around.
| Machine | CPU | GPU | Memory | Bandwidth | Fit budget used |
|---|---|---|---|---|---|
| Bosgame M5 (x1) planned | AMD Ryzen AI Max+ 395 (16C/32T) | Radeon 8060S iGPU (40 CU, RDNA 3.5) | 128 GB LPDDR5X unified | 256 GB/s on paper, ~210-220 measured | ~108-110 GB GPU-usable on Linux with GTT raised (96 GB cap on Windows). Fit budget used here: 108 GB. |
| Bosgame M5 (x2, RPC) planned | 2x Ryzen AI Max+ 395 | 2x Radeon 8060S over llama.cpp RPC (USB4 or 10GbE) | 2x128 GB | Per box; linking adds capacity, not speed (community: ~14-22% token-gen cost) | ~200-225 GB usable across two nodes. Fit budget used here: 210 GB; 210-225 GB = tight. |
| MacBook Pro (M4 Max) owned | Apple M4 Max, 14 cores (10P+4E) | 32-core GPU, Metal 4 | 36 GB unified | Apple spec for 14-core M4 Max | macOS caps GPU wired memory by default (~75% = ~27 GB). Raising iogpu.wired_limit_mb to ~32 GB is possible but leaves little for the OS. |
| Mac mini (M4) owned | Apple M4 | Apple M4 GPU (core count unknown) | 16 GB unified | Apple spec for M4 | ~2/3 of 16 GB usable by the GPU by default. Already busy as build box + Hermes/Ollama. |
| omarchy (AMD desktop) owned | AMD Ryzen 9 5950X | Radeon RX 6600 XT 8 GB (RDNA2, gfx1032) | Current RAM unknown (128 GB DDR4 upgrade planned) | GPU VRAM 256 GB/s; dual-channel DDR4 system RAM is far slower (~50 GB/s class, estimate) | 7 GB of the 8 GB VRAM for weights. Offload capacity depends on system RAM, which is unknown today; after the 128 GB upgrade ~110 GB of MoE experts could sit in RAM. |
| Ugreen DXP2800 NAS owned | Intel N100 (4 cores) | Intel UHD (iGPU) | 8 GB | Single-channel DDR5; not a model runner | Storage host. Already runs ollama/open-webui containers; only tiny models (e.g. Whisper on CPU) are realistic. |
| MacBook Neo owned | unknown | unknown | unknown | unknown | Specs unknown: machine was offline/unreachable on Nov 14 2026. |
Unknown specs stay unknown: the MacBook Neo was unreachable, and omarchy's current RAM isn't recorded (a 4×32 GB upgrade is planned). Those columns will change once the specs are read.
Matrix
| Model | Est. memory | Bosgame M5 (x1) | Bosgame M5 (x2, RPC) | MacBook Pro (M4 Max) | Mac mini (M4) | omarchy (AMD desktop) | Ugreen DXP2800 NAS | MacBook Neo |
|---|---|---|---|---|---|---|---|---|
| Qwen3.8-27B BF16 safetensors | ~62.4 GB | Quantize first | Quantize first | Quantize first | Quantize first | Quantize first | Quantize first | Unknown |
| Qwen3.8 Flash-Next BF16 safetensors | ~382.0 GB | Quantize first | Quantize first | Quantize first | Quantize first | Quantize first | Quantize first | Unknown |
| GLM-5.3-Flash safetensors (as published) | ~348.8 GB | Quantize first | Quantize first | Quantize first | Quantize first | Quantize first | Quantize first | Unknown |
| DeepSeek V4.1 Flash 1017 FP8/FP4 safetensors (ISTA card calls base FP8/FP | ~539.8 GB | Quantize first | Quantize first | Quantize first | Quantize first | Quantize first | Quantize first | Unknown |
| Qwen3.8-27B UD-Q8_K_XL GGUF + mmproj | ~38.0 GB | Fits ≈6.7 tok/s ceiling | Fits ≈6.7 tok/s ceiling | No | No | After 128 GB RAM | No | Unknown |
| Qwen3.8 Flash-Next UD-IQ4_XS GGUF (4 shards) + mmproj | ~103.3 GB | Fits ≈67.0 tok/s ceiling | Fits ≈67.0 tok/s ceiling | No | No | After 128 GB RAM | No | Unknown |
| GLM-5.3-Flash UD-IQ2_XXS GGUF + mmproj | ~112.2 GB | Tight / offload ≈37.2 tok/s ceiling | Fits ≈30.5 tok/s ceiling | No | No | After 128 GB RAM | No | Unknown |
| DeepSeek V4.1 Flash 1017 GSQ-RCO IQ1_M backbone + shared Engram IQ4_NL + | ~130.8 GB | No | Fits | No | No | No | No | Unknown |
| Qwen3.8-27B Mirai S 2.4-bit GGUF + mmproj + refusal control | ~16.7 GB | Fits | Fits | Fits | No | After 128 GB RAM | No | Unknown |
| Qwen3.8 Flash-Next GSQ-RCO IQ3_S GGUF + mmproj + refusal vector | ~92.8 GB | Fits ≈75.0 tok/s ceiling | Fits ≈75.0 tok/s ceiling | No | No | After 128 GB RAM | No | Unknown |
| GLM-5.3-Flash Maya-M IQ2_S GGUF + vision + refusal vector | ~127.0 GB | No | No | No | No | No | No | Unknown |
| Qwen3.8-27B Q4_K GGUF + mmproj | ~22.6 GB | Fits ≈12.3 tok/s ceiling | Fits ≈12.3 tok/s ceiling | Fits ≈23.4 tok/s ceiling | No | After 128 GB RAM | No | Unknown |
| Qwen3.8-27B BF16 GGUF + mmproj | ~62.4 GB | Fits ≈3.9 tok/s ceiling | Fits ≈3.9 tok/s ceiling | No | No | After 128 GB RAM | No | Unknown |
| Qwen3.8 Flash-Next UD-Q4_K_XL V2 GGUF (4 shards) + mmproj | ~123.1 GB | No | Fits ≈45.8 tok/s ceiling | No | No | No | No | Unknown |
| GLM-5.3-Flash UD-IQ2_XXS GGUF + mmproj | ~112.2 GB | Tight / offload ≈37.2 tok/s ceiling | Fits ≈30.5 tok/s ceiling | No | No | After 128 GB RAM | No | Unknown |
| DeepSeek V4.1 Flash 1017 UD-IQ1_M GGUF (5 shards) + mmproj | ~267.9 GB | No | No | No | No | No | No | Unknown |
| Whisper large-v3-turbo safetensors | ~5.6 GB | Fits | Fits | Fits | Fits | Fits | Tight / offload | Unknown |
| Qwen3-TTS 1.7B Base safetensors | ~8.5 GB | Fits | Fits | Fits | Fits | No | No | Unknown |
| Fish Audio S2 Pro safetensors | ~15.0 GB | Fits | Fits | Fits | No | No | No | Unknown |
| ACE-Step 1.5 safetensors (turbo DiT + LM) | ~8.0 GB | Fits | Fits | Fits | Fits | Fits | No | Unknown |
| YuE2 (ComfyUI repack) bf16 safetensors (ComfyUI) | ~17.1 GB | Fits | Fits | Fits | No | No | No | Unknown |
| Qwen-Image 2.1 safetensors (as published) | ~37.1 GB | Fits | Fits | No | No | No | No | Unknown |
| MOSS-Image Base safetensors (as published) | ~31.2 GB | Fits | Fits | Tight / offload | No | No | No | Unknown |
| Wan 2.2 T2V A14B safetensors (as published) | ~130.2 GB | Tight / offload | Tight / offload | No | No | No | No | Unknown |
| Wan 2.2 I2V A14B safetensors (as published) | ~130.2 GB | Tight / offload | Tight / offload | No | No | No | No | Unknown |
| HunyuanVideo 1.5 safetensors (as published) | unknown | Fits | Fits | Tight / offload | No | No | No | Unknown |
| FLUX.3 Klein (ComfyUI) unknown (gated) | unknown | Unknown | Unknown | Unknown | Unknown | Unknown | Unknown | Unknown |
| LTX-2.5 unknown (gated) | unknown | Unknown | Unknown | Unknown | Unknown | Unknown | Unknown | Unknown |
Method
- memory: Estimated memory = resident weight bytes x 1.05 + 4 GB (runtime + about 32K tokens of KV cache) for LLMs; media = files + 4 GB, an upper bound since ComfyUI/Diffusers offload components. Estimates, not measurements.
- tok_s: Token/s = memory bandwidth / (active params x bits-per-weight / 8). Bits-per-weight = GGUF bytes x 8 / GGUF parameter count. This is a theoretical ceiling for generation; real numbers are usually well below it. Two-box figures take off 18% for RPC only when the model does not fit one box (otherwise run it on one box). Not given where active params are unknown or the model needs conversion.
- verdicts: fits = fits in the GPU-usable budget; offload = needs raised limits or CPU/RAM offload; convert = original safetensors, quantize first; after-RAM-upgrade = MoE expert offload once omarchy has 128 GB; unknown = machine specs unknown.
Diffusion and speech models don't split across two boxes, so the M5 ×2 column for them means "on either box". Maya, Mirai and Strata builds use forked engines without llama.cpp RPC, so they get "No" on M5 ×2 unless they already fit one box.