Selection rationale · verified against live sources 2026-09-09

Models

The machine has ~110 GiB of realistically usable unified memory (128 GiB minus OS and KV-cache headroom). Every pick below was verified live against HuggingFace and the llama.cpp project — not from prior knowledge. Full research: research/models-deep-dive.md in the build docs repo.

Selected for this machine

ModelSize on diskWhy
Qwen3.8-27B (UD-Q4_K_XL + MTP draft + mmproj vision)~19.2 GBDaily driver. Vision + tool use + 262K context + draft-MTP speedup. Community: 26–32 t/s on this hardware. Downloading now (hotspot).
DeepSeek-V4-Flash (Strix-Halo-verified GGUF, 2.90bpw)~103 GBThe big frontier MoE. 1M context, MIT license, 90.8% API-parity verified, 20–27 t/s short context on this hardware. USB transfer pending.
Qwen3.8-Flash-Next (ROCmFP4 STRIX build, PLE streamed from NVMe)~113.5 GBPurpose-built Strix Halo quant — only 63–75 GiB memory-resident (51B PLE table streams from disk). 40–47 t/s on code with the HIP top-k fix that today's build includes. USB transfer pending.
GLM-5.3-Flash (6block IQ2_XS)~85 GBStaged for when llama.cpp support merges (3 PRs open, none merged as of 2026-09-09 — stock builds can't load it yet). 18B active → speed tolerates the aggressive quant. USB transfer pending.
Llama 3.3 70B (Q4_K_M)~40 GBRequested dense-70B baseline. Bandwidth-bound on this iGPU (~5–6 t/s est.) — benchmarked for comparison, not a serving model. USB transfer pending.

Considered and excluded — with reasons

Capabilities matrix (selected models)

CapabilityQwen3.8-27BDSV4-FlashFlash-NextGLM-5.3-Flash
Tool / function callingYes (jinja template)YesYesYes
VisionYes (mmproj-F16)Verify at loadYes (mmproj-BF16)Unverified for community quants
EmbeddingsSeparate embedding model (served on its own endpoint — see Architecture)
Draft-MTP speedupYes (Q4_0 draft, 1.3 GB)NoYes (Q4_K_M draft)PR-dependent
Context262K1M262K+1M

Why these quants

On a ~256 GB/s unified-memory iGPU, decode speed is memory-bandwidth-bound: bytes moved per token is the whole game, so the quality/size trade-off of the quant dominates throughput. Dense models (like the 70B) move every parameter every token and therefore crawl; MoE models (DSV4-Flash, Flash-Next, GLM) move only their active experts, which is why 100GB+ MoE models outperform a 40GB dense model here. The 27B dense at Q4_K_XL sits at a sweet spot: entire model + KV cache comfortably in memory, 26–32 t/s, full vision and tool support.