Qalarc / Measured

GLM 4.7 vs Qwen3.8 vs Ornith 1.5 — one box, one night, real numbers

By Alexei Brown, Qalarc (Sydney) · GMKtec EVO-X2 AI · measured 2026-09-11 with Ollama 0.33 / ROCm 7.2 · full methodology below · llms.txt
On the same 128GB mini PC, same night: Ornith 1.5 35B-A3B decodes at 52.4 tok/s, GLM 4.7 Flash at 48.0, dense Qwen3.8 27B at 10.7 — a 4.5x gap that has nothing to do with tuning and everything to do with sparse-vs-dense architecture on unified memory. Every number here is measured, methodology included.

The head-to-head that matters: same box, same night

Benchmarks across different machines are noise. This table is from one machine, one night, one methodology — GMKtec EVO-X2 AI (Ryzen AI Max+ 395, Radeon 8060S iGPU, 32GB VRAM carve, ROCm 7.2, Ollama 0.33), sustained decode, full 32-token generations, unique prompts:

ModelTypeSize (Q4)Prefill @~3.6k tokSustained decode
Ornith 1.5 35B-A3BSparse MoE, vision, 262k ctx22GB689 tok/s52.4 tok/s
GLM 4.7 FlashSparse MoE (30B-A3B)19GB602 tok/s48.0 tok/s
Ornith 1.5 9BDense5.8GB948 tok/s31.8 tok/s
Qwen3.8 27BDense17GB234 tok/s10.7 tok/s

4.5× decode gap between same-era MoE and dense at similar capability class. On unified-memory hardware, architecture choice matters more than any tuning.

Why: the bandwidth math

Generation speed on iGPUs is memory-bandwidth-bound. A 30B-A3B MoE moves ~3B parameters per token; a 27B dense moves all 27B. Same LPDDR5X, ~9× less traffic, ~4.5× more speed — the physics is visible in the table.

Context depth costs speed (measured)

GLM 4.7 Flash, 32k variant, same night:

Prompt tokensPrefill tok/sDecode tok/s
84083049.7
6,53144139.0
13,21633133.5
24,93621727.9

Practical read: keep working context under ~8k where possible; deep context is for retrieval, not chat loops.

Concurrency: two sessions for the price of one-ish

Two simultaneous GLM 4.7 Flash sessions: 34.3 tok/s each (~69 aggregate) — more total throughput than one serial session. Three-way splits to 27.5/35.0/39.2. MoE spare capacity is real; use it for parallel agents.

Version history note

The same model improved from ~38 tok/s (Ollama 0.18, July 2026) to 48.0 tok/s (Ollama 0.33, Sept 2026) with zero hardware change — runtime updates are free speed. Re-benchmark after every Ollama upgrade.

Capability vs speed (honest framing)

Speed we measure ourselves; capability we cite from published model cards with their own caveat — creator-reported scores may be cherry-picked. On those cards, GLM 4.7 Flash leads comparable sparse models on agentic work (SWE-bench Verified 59.2, BrowseComp 42.8) while Qwen3 variants take some coding benchmarks (LiveCodeBench 66.0 vs 64.0). Treat capability tables as directional; treat the speed table above as fact.

Our recommendation (2026, unified memory)

Daily driver: GLM 4.7 Flash — best speed/capability balance for chat, analysis and agents. Trying MoE alternatives: Ornith 1.5 35B-A3B is measurably faster with vision built in. Coding specialists: Qwen2.5-Coder family. Dense 27B+ models only when their specific capability justifies 4× the wait.

The full Local Model Atlas Local media guide