Use cases
codingagentic
Summary
User reports GLM-5.3-Flash at 74.0 tok/s for a single request on a Mac Studio M5 Ultra 256GB, with aggregate throughput rising to 116.6 tok/s at 4 concurrent requests using a patched oMLX.
Setup is oMLX 0.8.0 with a 4-bit abliterated MLX conversion of the 173 GB model, DFlash speculative decoding on (adaptive verify), several Claude Code and opencode terminals sharing one server.
Stock oMLX gave 76.1 / 70.9 / 91.7 tok/s at 1 / 2 / 4 requests; the fold patch plus batched kernel gave 74.0 / 84.3 / 116.6, a 27% gain at 4 requests. With speculation off, stock was 54.6 / 55.0 / 85.6 and patched 54.6 / 85.9 / 117.3. Gains only appear when requests overlap; a single request is not faster. The user attributes the ceiling to about 12.7 GB of weights read per step plus about 4.8 GB per extra request.
Compare this run
Medians of plain runs: one device, one request, no speculative decoding, the whole model in the device's memory. The report above is one setup, so read it beside them and not as one of them.
GLM-5.3 320B · 18B active at 4-bit by hardware
- 37.4 t/s · one run
Bars are medians on one scale, 0 to 37.4 tokens/s. Press ▶ to watch text arrive at a row's speed.
Models most run on the M5 Ultra 256GB
- 73.4 t/s · one run
- Gemma 4 31B16-bit13.9 t/s · one run
- 74.7 t/s · one run
- 17.0 t/s · one run
- 141.0 t/s · one run
- Qwen3.8 27B4-bit50.0 t/s · one run
Bars are medians on one scale, 0 to 141.0 tokens/s. Press ▶ to watch text arrive at a row's speed.
Everything reported on the M5 Ultra 256GB →Every GLM-5.3 report →Estimate GLM-5.3 on this card →