English documentation · runtime 7.24.4 · SDK 2.6.7. Content is maintained with runtime development; see each guide's scope and review date.
Tiel / Cyber-Tiel: AX Engine vs MTPLX (20 September 2026)
Status: Active Scope: public summary of the AX Engine four-machine native-API campaign Last reviewed: 2026-09-21 Owner: ax-code runtime
AX Code is the coding-agent harness. AX Engine is the Apple Silicon inference runtime it
bundles (signed 7.5.7 sidecar, ax-engine serve). This page records the current
matched Tiel peer numbers so the AX Code README does not keep the 19 September M3 Max
snapshot as if it were still the ranking.
Canonical tables, sample spread, TTFT, memory, and reproduction live in AX Engine:
Do not treat this page as a substitute for those artifacts.
What was measured
- Packs:
AutomatosX/AX-Tiel-Coder-35B-A3B-MLX-AXQ-MXFP4-MTP(AX Code managed default) andAutomatosX/AX-Cyber-Tiel-Coder-35B-A3B-MLX-AXQ-MXFP4-MTP(alternate). - Peer: MTPLX 2.11.3 sustained, same model files, same prompt token arrays.
- Primary metric: completion tokens/s including TTFT (native API entry to last callback).
- Decode tokens/s excludes the first callback. It is not the primary metric and is not AX Code session speed.
- Six measured samples per cell, cold KV, MTP depth 3, full-resident weights.
- Not default
ax-engine serve, not an AX Code coding session, not a quality claim.
Headline cells (M5 Max 128 GiB)
| Pack | Workload | Completion (AX vs MTPLX) | Decode (AX vs MTPLX) |
|---|---|---|---|
| Tiel (default) | python-lru | 194.88 vs 177.44 | 217.85 vs 198.40 |
| Cyber-Tiel (alternate) | python-lru | 219.43 vs 194.15 | 249.01 vs 219.84 |
194.88 is the public number for the managed default pack. 249.01 is the fastest decode cell in the campaign; it is Cyber-Tiel, not the default, and managed read-task probes for that pack remain unresolved.
On Mac mini M4 Pro 64 GiB, Tiel python-lru completion is 90.46 vs MTPLX 92.23.
AX Code recommends an M4 Pro with 48 GB or more for the 35B pack; M5 Max 128 GiB is
the campaign host, not the minimum. AX Code itself still runs from 8 GB on M2 with cloud
or lighter local models.
Do not mix with Qwen 3.8 27B
AX Engine’s first-run pack is dense Qwen 3.8 27B AXQ 6-bit MTP. That pack streams 20.84 GB per token and is already at the DRAM ceiling in direct AR. Product-path MTP is 31.05 tok/s decode on Mac mini M4 Pro 64 GB (2.43× mlx-lm 12.78) and 76.90 tok/s decode / 795.3 tok/s prefill on M5 Max 128 GB (2.76× mlx-lm 27.90). It is not in AX Code’s managed catalog. Do not quote 76.90 next to Tiel 194.88 as if they were the same metric or the same model. See AX Engine Qwen 27B performance.
Historical 19 September M3 Max snapshot
Tiel prefill/decode benchmark remains on disk as a same-host native-phase snapshot (M3 Max 128 GiB, engine build identified as 7.4.0, median of three, AX 47–58 decode vs MTPLX 92–96). That ranking does not supersede the 20 September four-machine campaign. Do not mix the two tables.
Session speed
Short-prompt native decode does not establish a floor for a coding session with tens of thousands of context tokens, tools, or HTTP sidecar transport. See Interpreting local response speed and the AX Code / OpenCode client retest (Qwen 3.8, a different model).