Test record: I upgraded the EVO-X3 production control plane from a locally patched Lemonade 11.8.0 daemon to the official Lemonade 11.9.0 release. The model files, inference binaries, launch arguments, aliases and runtime state stayed fixed. Executable accuracy moved from 30/40 to 31/40, fixed prefill changed by -0.04%, fixed decode by -1.93% and streaming TTFT by -2.60%. I retained 11.9 in production.

This was deliberately not an inference-engine upgrade. Lemonade owns the API, model registry, routing and lifecycle on this machine; custom backends still do the token work. Keeping that boundary fixed made the question answerable: does the new control plane preserve the service I already qualified?

It did. The interesting result is operational rather than dramatic. Official 11.9 absorbs the long-stream timeout behaviour I previously carried as a local patch, restores the same routes and produces performance inside the recent 11.8 range.

The paired benchmark data and upgrade manifest contain the compact figures and identities.

What 11.9 changes

The release’s most useful change for this system is at commit bb39eafc: streaming silence now follows the configurable global_timeout rather than a fixed 120-second limit. A genuinely large Qwen prefill can be quiet for longer than two minutes, so the old constant could kill a healthy request. Production keeps its four-hour bound.

The 11.9 release notes also record:

  • corrected remote-origin handling and a new allowed_origins setting;
  • recovery for cache and models misplaced by the 11.8 FHS migration;
  • an experimental AMD HRX backend for gfx1100/gfx1151 Linux;
  • llama.cpp b10723 with Qwen3-Next support;
  • per-file typing for models in reserved extra-model directories.

I did not enable HRX or switch to packaged b10723. The latter is important because b10723 is also the build family in the current Strix Halo lazy-loading regression. The custom dispatch boundary kept that unrelated inference change out of this upgrade.

The locked before-and-after run

Both arms used benchmark signature 16cf9033 and model fingerprint f425f930. The test is a fixed 40-task HumanEval/MBPP subset at temperature zero and seed one, with generated Python executed in a networkless container. The Qwen profile remained 524,288 total context, two slots, 40 Vulkan GPU layers, Q8 K/V and a 4 GiB host prompt cache.

Metric 11.8 before 11.9 after Change
Executable tasks 30/40 31/40 +1 task
Task errors 0 0 unchanged
Fixed uncached prefill 195.661 tokens/s 195.579 tokens/s -0.04%
Fixed decode 18.931 tokens/s 18.567 tokens/s -1.93%
Median streaming TTFT 0.2624 s 0.2556 s -2.60%
Median task wall 7.237 s 7.627 s +5.39%
P95 task wall 41.654 s 42.173 s +1.25%

Thirty-seven of forty extracted programs were byte-identical. Three changed, and one of those changed its executable outcome from fail to pass. I do not attribute that improvement to a smarter Lemonade release: the model and sampler were fixed, but tiny runtime variation can still change an extracted program at the edge of a decision.

The performance deltas also sit inside recent 11.8 history. Fixed decode had ranged from 17.927 to 18.881 tokens/s and prefill from 189.580 to 197.334. On that evidence, -1.93% is run noise rather than a material regression.

Five routes, the same resolved targets

The capability sweep verified model selection and included switch costs where applicable:

Route Target 11.8 11.9 Result
Jarvis Qwen3.5 4B, pinned NPU 2.080 s 1.818 s same marker
llm Qwen3.8 Flash Next 125B 0.846 s 1.908 s same marker
llm-reasoning Ornith 1.5 35B 17.587 s 15.878 s same marker
deepseek-code DeepSeek Coder V2 47.596 s 42.249 s same marker
deepseek-flash DeepSeek V4 Flash 76.847 s 53.724 s same marker

I would not read five independent speed claims from this table. It includes cold model selection, storage state and cache effects. The fixed repeated probes above are the controlled performance comparison; this table proves that the routes still resolve, launch and answer.

One restart race was worth keeping

During the first production restart, the Qwen wrapper’s memory preflight rejected an early concurrent warmer attempt. Lemonade evicted, retried and reached the correct residency. Both benchmark and capability suites then passed.

I changed the retained deployment script so it restores the pinned NPU helper first and starts the Qwen warmer second. That makes a future replay deterministic without weakening the memory guard.

The final service reports the official 11.9 binary, thirteen unchanged model IDs, six unchanged aliases and byte-identical state JSON. LAN and Tailscale listeners return version 11.9.0. No failed user unit or matching AMD GPU fault was recorded. The unpinned Qwen model and pinned NPU helper are both healthy.

Where I would stop claiming

This test qualifies Lemonade as the control plane around my existing backends. It does not benchmark HRX, packaged b10723, a new model or the release on a default installation. The route timings are operational smokes, not clean engine comparisons.

The decision is still clear: retain official Lemonade 11.9. It removes one local patch, preserves the measured service and leaves a complete 11.8 rollback available. That is a worthwhile upgrade even without a tokens-per-second headline.

Tested 2 September 2026 on GMKtec EVO-X3, Ryzen AI MAX+ 395 / Radeon 8060S, 128 GiB unified memory. Lemonade 11.9.0 artifact SHA-256 a18beaf7…eb204d; installed lemond SHA-256 0a164f82…bcb118.