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:
| Model | Type | Size (Q4) | Prefill @~3.6k tok | Sustained decode |
|---|---|---|---|---|
| Ornith 1.5 35B-A3B | Sparse MoE, vision, 262k ctx | 22GB | 689 tok/s | 52.4 tok/s |
| GLM 4.7 Flash | Sparse MoE (30B-A3B) | 19GB | 602 tok/s | 48.0 tok/s |
| Ornith 1.5 9B | Dense | 5.8GB | 948 tok/s | 31.8 tok/s |
| Qwen3.8 27B | Dense | 17GB | 234 tok/s | 10.7 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.
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.
GLM 4.7 Flash, 32k variant, same night:
| Prompt tokens | Prefill tok/s | Decode tok/s |
|---|---|---|
| 840 | 830 | 49.7 |
| 6,531 | 441 | 39.0 |
| 13,216 | 331 | 33.5 |
| 24,936 | 217 | 27.9 |
Practical read: keep working context under ~8k where possible; deep context is for retrieval, not chat loops.
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.
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.
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.
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