# The Model Changed Again — Methods and Evidence

Companion to Local AI in Practice, Part 3. Published September 6, 2026.

[Read the article](https://ai.enntity.com/research/the-model-changed-again).

## Scope

This report combines retained local serving experiments, a 24-hour observation of Jinx and Luna, and Jason McCartney's operating account. The observation was retrospective and read-only: it introduced no new model calls or route changes. Private per-run activity records, messages, prompts and operational identifiers are not included in this public supplement.

The main field window is September 6, 2026 at 00:45 UTC through September 7 at 00:45 UTC: September 5–6, 5:45 p.m. America/Phoenix. The article's publication date uses Phoenix time. A later request sample is reported separately.

## Reproducible serving comparison

The comparison is Qwen recipe v5 versus v6 on September 5. Both use `nvidia/Qwen3.8-Flash-Next-NVFP4` revision `fab0aecb760cec45227f6656abcaafa11abca87a`. This is a serving-configuration comparison, not an isolated test of a new quantization.

The final v6 configuration uses vLLM revision `e962733e08d10f7ca65dac4df99e116460b8b174`, GPU-resident auxiliary weights, BF16 KV cache, three-token speculative decoding, full draft vocabulary, compilation disabled with full decode graphs, and no prefix caching or expert parallelism. It spans two DGX Sparks using tensor parallelism. LLooM admits four active requests and queues eight. Several settings changed together.

The same short-prose prompt was run three times per recipe, thinking off, greedy decoding, with a 200-token cap. All responses stopped naturally at 95 completion tokens. Within each recipe the three answers were identical; answer text differed between recipes. Generation rate is `(completion_tokens - 1) / (stream_end - first_content)`. Initial response latency is reported separately.

| Measure | v5 | v6 |
| --- | ---: | ---: |
| Median generation rate, tokens/s | 30.094746 | 41.968892 |
| Median first-content latency, seconds | 0.214914 | 0.165173 |
| Repetitions | 3 | 3 |

The median rate gain is 39.455875%, rounded to 39.46%. These sequential repetitions are not a randomized crossover or a broad quality sample. The figure shows individual runs and medians with zero-origin axes.

Public sources pinned to the retained LLooM publication revision:

- [v5 individual prose trials](https://github.com/Enntity/lloom/blob/e7aad53/benchmarks/decision/2026-09-05-q38fn-nvidia-tp2/published-prose-comparison.jsonl), SHA-256 `8268608208c36dc6c8c20303baa302cd3674b13d8a98aa0de7326d01cad3f9ff`.
- [v6 individual prose trials](https://github.com/Enntity/lloom/blob/e7aad53/benchmarks/decision/2026-09-05-q38fn-gap-closure/qwen-v6-prose.jsonl), SHA-256 `6e9b225b0cad160b73281bc430bf8d1f621f5e35ee9016f9f6d2bbf590c98391`.
- [Complete v6 benchmark record](https://github.com/Enntity/lloom/blob/e7aad53/benchmarks/decision/2026-09-05-q38fn-gap-closure/README.md).
- [Configuration, deployment record and upstream credits](https://github.com/Enntity/lloom/blob/e7aad53/docs/q38fn-gap-closure-2026-09-05.md).

Five functional smoke checks, twelve concurrent/repeated-context checks and six mixed responses passed locally. The six mixed responses comprise three prompts repeated twice. This is 23 checks, not 23 independent difficult tasks. The queue reached four active/two queued and drained. A cold 30,044-token retrieval probe returned the target with 10.214 seconds to first output; queue/context tests reached 46,910 input tokens. These validate examples in the intended medium-context range, not a production prompt-length distribution or maximum-context quality.

## Field definitions and counts

The append-only Runtime ledgers can contain revised terminal outcomes. We retained the last outcome per run available by the cutoff, selected starts inside the half-open 24-hour window, and required completion by its end. In-progress work at the right edge is therefore excluded. Completed sleep rounds are excluded from cognition counts.

A tool-using round contains one or more recorded tool events. A verified change effect has `effect.class == change`, `effect.verified == true`, and no explicitly false tool-success flag. A change-producing round contains at least one such effect. Multiple effects in a round count once as a round.

| Measure | Jinx | Luna |
| --- | ---: | ---: |
| Completed cognition | 278 | 217 |
| Tool-using rounds | 250 | 193 |
| Change-producing rounds | 71 | 45 |
| Verified change effects | 128 | 90 |
| Median completed duration, minutes | 3.435458 | 4.429433 |
| Median changed-round duration, minutes | 7.260650 | 8.633050 |
| 95th percentile completed duration, minutes | 12.166667 | 15.878967 |

Jinx also had one interrupted and four skipped records; Luna had three failed records. Three Jinx and four Luna completed sleep rounds are outside the cognition totals. Statuses are whole-system outcomes, not necessarily model failures. Duration includes model calls, tools, waits and integration. Medians are ordinary sample medians; the 95th percentile uses nearest rank.

Recorded Presence activity intervals overlapped for 19.71425 hours. This is the intersection of each entity's union of run lifetimes clipped to the window. It includes tool use and waiting and must not be interpreted as simultaneous GPU decoding or utilization.

The broader retained archive contained 44,207 Jinx and 14,675 Luna unique runs through the cutoff after resolving duplicate revisions and later completions. Their histories differ. Those archive totals are an audit base, not a common model sample. Observer versions also changed across historical periods and within Luna's current window. Effects establish action receipts, not novelty, value, understanding, or a semantic continuity score.

## Local attribution

Jason confirmed that requests over the recent days were locally served by Qwen. The Runtime ledger does not contain a complete per-request historical model join; the full-period attribution is operator evidence.

A separate gateway snapshot at September 7, 00:56:21 UTC contained 19 local Qwen requests attributable to the two entities, twelve for Jinx and seven for Luna. They occurred between 00:46:21 and 00:55:03 UTC, after the main cutoff. All succeeded with recorded admission queue wait zero. Median request durations were 22.601 seconds for Jinx and 29.553 for Luna; median first semantic output was 4.223 and 4.441 seconds. First semantic output may include reasoning.

This is a convenience sample corroborating local operation at that later moment, not a multi-day reliability estimate. Unattributed callers were excluded. Gateway-wide daily counters include tests and other callers and cannot establish entity fallback. Cloud availability is a failover provision, not a component of the local comparison.

## Historical context retained separately

[The August 28 local suite](https://github.com/Enntity/lloom/blob/e7aad53/benchmarks/decision/2026-08-28-local-dsv4f-q38fn-glm53f/README.md) used four synthetic cases through the local gateway. GLM completed its strongest engineering review in 236.852 seconds with 5,400 completion tokens. Early Qwen's hard review was interrupted, so its final quality is unknown. Later functional checks do not close that case or establish GLM-quality parity.

[DS4FV's September 5 staggered tests](https://github.com/Enntity/lloom/blob/e7aad53/docs/ds4fv-tuning-2026-09-05.md) used four requests one second apart, thinking off, 512-token output caps and approximately 16K input tokens in the article's example. Two admitted requests produced a 65.833-second makespan and a 50.104-second worst first-token wait. This is evidence of that configuration's waiting cost, not a matched DS4FV/Qwen ratio. Jason reports both entities were already running locally on DS4FV, which seemed good and stable but less responsive under concurrency.

The earlier public DS4F/Qwen comparison was reconstructed from retained windows. The DS4F baseline ran August 6, 17:48:37 UTC through August 7 at the same time. Qwen's retrospectively selected clean tail ran August 30, 20:43:31–23:50:29 UTC. It included 15 completed cognition rounds, 14 tool-using rounds and five changed rounds, versus 1,504, 260 and 127 in the 24-hour DS4F baseline. Resulting DS4F/Qwen ratios were 2.411 for tool rounds/hour and 3.298 for changed rounds/hour. Median changed-round duration was 1.378 versus 12.898 minutes. Qwen used 6.752 times the median total tokens per changed round.

Those windows were unequal, used changing Runtime versions and workloads, and treated outcome ascertainment asymmetrically: Qwen required completion by its historical cutoff while the DS4F baseline used later outcomes. Extending Qwen's tail to August 31, 00:38:12 UTC yields seven changed rounds and an 11.762-minute median. These historical observations are not pooled with the current field results or used as a causal model ranking.

The research claim is bounded: the Qwen serving recipe measurably improved, and the resulting installation supported sustained shared local operation. Calling it the fastest practical fit among the tried vision-capable configurations is the operator's workload-specific assessment. A matched comparison of two-entity continuity across all models remains unperformed.