Strata for
local AI

RenderMasterBot's core value is local AI-powered graphics assistance: supplied scene context and inference requests stay on the workstation instead of going to an external cloud-model service. Strata extends the investigation to a larger local model for questions that can tolerate offline latency.

The intended use is a two-track assistant: native models for interactive work, and a stronger offline candidate for difficult code and evidence-based engineering questions. These experiments test that candidate; Strata is not yet the production Unreal route, and no default model was switched.

The question is therefore useful capability—not parameter count alone. Can the configuration deliver a correct answer within this workstation's resource constraints? The results below separate speed, deterministic checks, preliminary QA, and incomplete deliveries.

These compare configured systems, not isolated runtime or quantization effects. Model families, weights, KV policies, templates, memory modes and study timing differ.

The original v3 code suite had four authored fixtures: ring-buffer repair, GPU interval/percentile calculations, a bounded logical-proposal validator, and a bilingual capture-status formatter. A whole case passes only when its required checks and frozen interface constraints pass.

QA is a preliminary blind agent rubric on development fixtures, one reviewer per answer—not calibrated human judgment or an independent general-intelligence benchmark. IQ2 and Q4 have no formal QA ranking; later tails did not rerun their original QA/code cells.

Each row is one independently qualified study/candidate/case group: three repetitions of a 512-token text workload. Thinking was disabled where supported; GPT-OSS used low reasoning rather than no reasoning. Values are API-reported decode medians; ranges show all three observations. Intentional length termination is expected here.

The two panels deliberately use different scales. IQ2's 21.4 and Q4's 3.9 tokens/s belong only to their later sealed tails—not to repaired original runs. No p95, confidence interval or pooled cross-study ranking is claimed.

The tested policy required at least 4 GiB sampled available physical memory and 2 GiB sampled available system commit. System commit tracks allocation headroom, so it can approach its floor even when physical RAM looks comfortable.

These are whole-system observations, not model-exclusive allocation or exact transient peaks. Requested expert budgets and page-release hints are not RAM caps. Earlier failed buffered and mapped Q4 policies remain in the history below.

1. Original IQ2

   Stopped after three records on malformed structured output. The interrupted missing-fourteen supplement later saved nine responses and one pending intent; its lifecycle was never sealed, so those nine responses remain diagnostic.

2. Original Q4

   Stopped after eight records on its owned deadline. Its fourth original QA case timed out at 1,800 seconds. Three observed code-case passes do not turn this failed segment into a formally accepted result.

3. Independent recovery

   Q4 tail9 and IQ2 tail4 sealed 13 fresh records in their own scopes: seven complete context/workflow responses and six capped speed samples. Complete delivery is not a passed correctness check. They did not repeat the failed original cells, the interrupted IQ2 records, or the pending intent.

The original full-matrix success flag remains false. These observations do not establish OOM, a network fault, the exact cause of an interruption, or overall model inability.

Strata is the inference runtime in this investigation. Its tested model was Qwen3.8-Flash-Next, in IQ2_XS, IQ3_XXS, IQ3_S and UD-Q4_K_XL configurations. The native comparison used Ollama with GPT-OSS 20B, Qwen3.5 9B and Qwen3-VL 8B.

The experiments separated installation, delivery, deterministic checks, preliminary QA and resource qualification. The tested configuration had to deliver a useful, correct engineering answer within the machine's resource and lifecycle constraints.

Strata's upstream design places frequently used sparse-expert weights on the GPU and uses host-memory or storage tiers for the rest. That makes a larger model accessible without fitting every weight into VRAM; it does not remove host-memory pressure or guarantee interactive latency. Upstream Strata project ↗

My takeaway: local deployment is feasible, but a useful engineering assistant needs correctness, resource safety, reliable lifecycle handling and human acceptance—not just a larger model or a valid JSON response.

The chronology includes installations, planning, successful narrow tests, failed gates, interrupted runs and independently sealed tails. Open any entry for its observed result and specific limits.

1. H012026-10-0817:58 UTCInitial load and structured-output checkcompleted within narrow scope+

   deployment check

   IQ2_XS loaded in 42.25 seconds; a supplied strict-schema probe returned in 2.17 seconds and passed validation. The model had a configured 65,536-token context, but full-context stress was not run.

   Limits Structured output was prompt-and-validate rather than constrained decoding. This tested the runtime and one contract, not live Unreal work or answer expertise.

2. H022026-10-0817:58 UTCFirst four-case Unreal-assistance evaluationrequests completed; preliminary quality failed+

   quality

   Four complete answers and confirmed unload. Preliminary review scored 27/40 against a frozen 34/40 threshold. Scope and structured-plan cases were stronger; profiler interpretation and experimental design contained substantive errors.

   Limits Four development fixtures, no live engine access. Review was agent preliminary judgment; human acceptance remains pending.

3. H032026-10-0818:00 UTCHistorical-profiler demonstration taskone response completed; acceptance pending+

   demonstration

   One supplied-evidence task returned a complete answer in 98.61 seconds. The input profiler record was a historical September 17 capture, not a new measurement.

   Limits No diagnosis was verified through new profiling or an executed optimization. This is a demonstration separate from the four-case suite.

4. H042026-10-0818:17 UTCFirst configured-system comparison attemptaborted before Strata phase; cleanup gate failed+

   comparison attempt

   The planned 16-record comparison saved eight Ollama answers, then stopped because model release was not confirmed. No Strata response belongs to this attempt.

   Limits Preserved diagnostic attempt. It is not a completed two-backend comparison and must not be merged with the later replacement protocol.

5. H052026-10-0818:23 UTCRevised Qwen3.5 versus Strata comparison16 main responses completed; full content acceptance failed+

   configured-system comparison

   Four tasks, two rounds per backend, plus a separate strict-schema request per backend. Preliminary totals were Ollama 23.5/40 in both rounds and Strata 29/40 then 32/40. All main rounds failed full acceptance. Median preloaded nonstreaming request times were 4.23 seconds and 7.54 seconds respectively.

   Limits The directory calls the fixture simulated, but these were local model requests over simulated/supplied tasks, not mocked inference. No live Unreal operations. Differences cannot be assigned solely to runtime, parameter count, quantization or reasoning.

6. H062026-10-0818:48 UTCGPT-OSS low-effort follow-upeight main responses completed; preliminary quality failed+

   historical comparison follow-up

   GPT-OSS 20B completed eight low-effort requests. Preliminary totals were 24/40 and 25/40; neither round passed acceptance. The run compared against earlier Qwen/Strata observations rather than rerunning a randomized three-way benchmark.

   Limits Reviewer variation was not calibrated. The first question included extra initialization, and the full timing aggregate retained it. A separate schema request is not folded into the main quality/timing total.

7. H072026-10-0818:51 UTCGPT-OSS high-effort diagnosticfour calls recorded; all final answers truncated+

   reasoning diagnostic

   One four-case round used high reasoning and doubled output caps. All four responses ended at the length limit; zero complete final answers were delivered.

   Limits Reasoning effort and cap changed together. No full high-effort quality score was assigned, and this track cannot replace the low-effort main track.

8. F012026-10-08UTC · date onlyHigher-quantization feasibility assessmentassessment completed; predictions not measurements+

   read-only planning

   Reviewed IQ3_XXS, IQ3_S, UD-IQ4_XS, UD-Q4_K_XL and UD-Q6_K_XL against historical RAM, commit, disk and documented runtime requirements. IQ3_XXS was the first bounded trial; higher formats were initially untested and Q6 required separate compatibility work.

   Limits Download sizes and rough RAM estimates were planning figures, not measured candidate peaks. IQ4 and Q6 have no completed local evaluation in this evidence set. Later IQ3_S/Q4 work does not retroactively change this assessment.

9. H082026-10-0819:06 UTCIQ3_XXS installationinstalled and verified; not automatically activated+

   installation

   The higher-weight option was installed beside IQ2_XS, with recorded file verification and original weights preserved.

   Limits Installation is not an answer-quality improvement, memory qualification or default-provider replacement.

10. H092026-10-0819:30 UTCTwenty-call quantization, budget and grounding study20 calls completed; proven quality upgrade not established+

    controlled quality study

    Separate tracks compared matched 32K IQ2/IQ3_XXS profiles, wider reasoning budgets, an evidence policy with and without supplied grounding, and two independent holdout cases. Matched original cases scored IQ2 28.5/40 versus IQ3 26/40; wide hard cases 12/20 versus 11/20; holdouts 19.5/20 versus 18/20. Only 5/20 answers passed the full preliminary content contract; all six structured calls passed schema validation.

    Limits The new 32K resident IQ2 control was not the original 64K launcher. No rescore or winner-selected rerun. Unguarded cold loads sampled physical headroom as low as 0.02 GiB for IQ2 and 1.24 GiB for IQ3, so these comparison profiles were not daily-use recommendations.

11. H102026-10-0819:50 UTCGuarded IQ3_XXS memory qualificationqualified for two tested workloads; quality not scored+

    resource-only qualification

    Two separate supplied synthetic holdout calls completed in 84.16 and 63.88 seconds. Sampled whole-system RAM peaked at approximately 33.10 and 33.14 GiB. A guarded optional profile reduced observed memory demand at a throughput cost; the existing default stayed unchanged.

    Limits Two workloads do not establish broad reliability. Allocation headroom, expert budgets and file-release hints are not a production RAM cap; one-second sampling can miss transient peaks.

12. H112026-10-0822:21 UTCFirst IQ3_S installation attemptpartial/unsealed; superseded by later verification+

    installation attempt

    The saved progress record showed approximately 46.0 GB of present final-or-partial data against an 83.6 GB weight total. No installation summary or verification terminal is present for this directory.

    Limits Present/downloaded bytes are not a completed publisher-hash check. Saved artifacts do not establish the exact reason this attempt ended.

13. H122026-10-0822:44 UTCSecond IQ3_S setup attemptsetup validator failed; later verification separate+

    installation attempt

    The official setup returned success, but the wrapper recorded installation failure because tokenizer/template artifacts were not found by its check. Protected assets remained unchanged, and models were not run.

    Limits This failure belongs to setup validation; it is not a model-quality or inference-speed failure.

14. H132026-10-0822:54 UTCExisting IQ3_S pack verificationverified existing files; no inference+

    installation recovery verification

    Verified the already-present weights and native tokenizer/template pack without downloading or rerunning official setup. IQ3_S was not activated as default.

    Limits Quality and resources were not tested by this verification. It supersedes the missing-pack installation diagnosis, not its historical record.

15. H142026-10-0822:55 UTCUD-Q4_K_XL installationinstalled and pack-verified; no inference+

    installation

    Installed and verified the Q4 files/native pack alongside existing options. Protected assets remained unchanged; the default profile was not activated or replaced.

    Limits Different file-tier I/O and compatibility conversion make later Q4 trials configured-profile tests, not pure quantization isolation.

16. H152026-10-0823:21 UTCHigher-precision coding and QA suite attempthalted after one IQ2 call by resource gate+

    aborted quality/resource study

    A planned four-quant suite included four coding fixtures and four QA fixtures per model. Only one IQ2_XS generation call was entered before resource qualification failed; sampled system-commit headroom fell to 0.35 GiB. Later candidate calls were not run.

    Limits Not a completed four-model coding benchmark. Unattempted cells are absent evidence, not failed answers; the stop does not prove OOM or a candidate-quality disadvantage.

17. H162026-10-0900:08 UTCBuffered-cache performance qualificationIQ3_S qualified; Q4 safety gate failed+

    bounded throughput/resource probe

    Two 512-token, no-thinking calls were attempted. IQ3_S completed its phase; Q4 failed the minimum sampled physical-headroom policy, with a 3.74 GiB minimum below the 4 GiB gate. Later samples were stopped.

    Limits Intentional length termination is a bounded throughput observation, not a completed high-quality developer answer. No quality evaluation.

18. H172026-10-0900:14 UTCMapped-page-release performance probeIQ3_S qualified; Q4 safety gate failed+

    bounded memory-policy probe

    A distinct mapped-page-release tuning study attempted two calls. IQ3_S completed; Q4 again crossed the physical-headroom gate, at 3.91 GiB. This failed configuration remains recorded.

    Limits Not pooled with the first probe. A page-release hint does not cap total process or system memory.

19. H182026-10-0900:20 UTCDirect-NVMe pageable versus pinned expert budgetsfour phases completed and resource-qualified+

    four-profile bounded speed study

    One no-thinking 512-token observation per profile: IQ3_S pageable24 18.2 tokens/s, pinned16 20.4 tokens/s; Q4 pageable24 5.9 tokens/s, pinned16 6.7 tokens/s. Actual expert residency was below requested budgets. The profiles used direct file reads rather than the earlier buffered-cache setup.

    Limits Single observations, fixed order, shared OS cache, intentional cap termination. These are historical API decode rates for that exact profile/workload, not final-answer latency, statistical superiority or globally optimal settings.

20. H192026-10-0900:46 UTCLong-reasoning IQ3_S validator exposureone long request passed its recorded exposure/resource/code gates+

    stability and bounded code contract

    One high-reasoning request used the pinned16 profile, produced 5,700 generated tokens over 470.27 seconds, and exposed nonempty reasoning. The independent bounded logical-proposal validator passed 54/54 checks. The sampling coverage gate passed with 443 call-time samples.

    Limits Reasoning characters are not an exact reasoning-token count. One validator request does not establish full developer quality, production stability or live Unreal automation.

21. H202026-10-0902:08 UTCAll-installed-options v1failed before benchmark calls; superseded+

    broad matrix attempt

    The 179-request matrix planned ten artifact/configuration options. No generation or embedding calls were entered. Preloads failed, an installed-asset/protected-state gate failed, and the study halted before Strata phases.

    Limits Zero benchmark responses: no quality or speed ranking can be drawn. The installed-asset check does not by itself establish the causal mechanism of the drift.

22. H212026-10-0902:30 UTCPrivate-alias readiness isolationpassed; zero inference+

    metadata readiness check

    Verified isolated alias bindings using metadata GET/show checks only. No model loads, generations or embeddings were performed; protected state and released ports passed the recorded readiness checks.

    Limits Readiness is not model-performance or capability evidence. GGML and llamacpp names identify installed artifacts; their native GGUF route is not evidence of two independently implemented engines.

23. H222026-10-0902:32 UTCAll-installed-options private v2interrupted/unsealed; superseded by checkpoint v3+

    broad matrix attempt

    The saved summary counted three entered generation calls, no embeddings, and a pending call in the first GPT-OSS phase. It remained marked running and incomplete without a final terminal.

    Limits A persisted pending call is not verified dispatch or successful delivery. No lifecycle success is invented from later idle-state observations.

24. H232026-10-0919:21 UTCDetached-launch survival probelaunch refused before inference+

    runner infrastructure probe

    A newly created child remained inside a Windows Job, so the launcher refused detached execution and stopped its owned child. The evidence explicitly says inference had not started.

    Limits An execution-lifetime infrastructure failure, not a model-quality or speed result. The directory timestamp differs from the actual recorded event time; the latter is used here.

25. H242026-10-0919:28 UTCCheckpointed all-installed-options v3controller finished; successful full matrix remained incomplete+

    checkpointed engineering matrix

    Canonical attempts reached 156/179: 151 generation calls and five embeddings. Eight candidate segments completed; original IQ2_XS failed after three records on malformed structured output/response contract, and original Q4 failed after eight records on an owned deadline. One earlier Q4 interruption attempt with two entered calls remains separately excluded. Synthetic tracks covered code, QA, context, workflow, text throughput, assisted vision for eligible models, and embeddings.

    Limits execution_state complete does not mean benchmark_execution_complete. Failed original candidates are not promoted by later supplements. No pooled all-time ranking.

26. H252026-10-1004:06 UTCIndependent IQ2 missing-fourteen-cell supplementinterrupted/unfinalized; diagnostic only+

    supplement attempt

    Planned fourteen missing cells without altering v3. Nine complete response records were persisted; the tenth entered-call counter was a pending intent with no finalized response. The historical lifecycle/release remained unverified.

    Limits All nine records are diagnostic only, never formal quality or speed samples. The pending cell is not confirmed dispatched. Four never-attempted tail targets were later tested separately.

27. H262026-10-1022:15 UTCIndependent Q4 nine-cell tailnine requests attempted and terminal-sealed; six complete context/workflow responses+

    sealed supplement

    Completed the nine-cell tail without importing original Q4 measurements. Six complete answer deliveries and three intentional length-bounded speed samples were recorded. Its own three-repetition text-decode group had median 3.9 tokens/s, range 3.8–4.1.

    Limits Original Q4 deadline/lifecycle failure remains a failure. This completed tail is not a completed full Q4 benchmark or a re-run of its failed original cells.

28. H272026-10-1022:42 UTCIndependent IQ2 four-cell recovery tailfour requests attempted and terminal-sealed; one complete workflow response+

    sealed recovery segment

    Tested only four never-attempted targets: one workflow case and three speed repetitions. It did not repeat the original failed cell, interrupted nine responses, or pending intent. Its own text-decode group had median 21.4 tokens/s, range 18.9–22.0.

    Limits The original IQ2 malformed-response failure and interrupted lifecycle remain unchanged. Four scoped tail calls are not a repaired full IQ2 benchmark.

The five notes below cover runtime engineering, mock regressions, evolving review snapshots, interrupted-byte recovery and the final evidence snapshot. They are supporting records, not additional model runs.

Public exports omit raw answer/reasoning transcripts, private machine paths, runtime commands, local endpoints, process identifiers, model weights and secrets. Publication organized existing evidence only: no model evaluations rerun, new model calls, live-state checks or weight rehashes.