The user requires at least 128K context, complete AWQ/NVFP4 weights, and vision.
The GPU-only 128K profile at memory fraction 0.95 failed before weights loaded: CUDA free memory was 30.2 GiB of 31.82 GiB, below the requested 30.23 GiB. This is a device-reservation failure, not a model- or KV-fit result.
The earlier no-MTP selective-UVA branch passed functional, vision, and a bounded context scout but failed its repeated quality gate at 27/30. Its failure remains retained under native/candidate/quality.json; it is not combined with evidence from the current MTP branch.
The current MTP-UVA3 branch uses target Triton attention and draft attention, BF16 KV, default FP32 SSM, MTP3, memory fraction 0.93, and generic 3 GiB UVA offload. Startup records 19.11 GiB GPU weights, 3.02 GiB actual generic UVA offload including 0.86 GiB visual offload, 9.7 GiB KV, and 138519 cache tokens. Functional evidence is 27/27, vision evidence is 18/18, and enabled-thinking reasoning evidence is 9/9.
The completed quality scout is 10/10 for one repetition, with all finish reasons stop. Its computer-science item returned FINAL=B, using 5217 completion tokens and 5058 reasoning tokens. The completed repeated quality diagnostic then passed 30/30, with all finish reasons stop. The 4-request strict capacity scout passed 4/4 with unique cache and 32-word strict output. The completed strict 100-request C1 short-output capacity population passed 100/100 with zero errors and all requests eligible: decode p50 51.88547695919587 tok/s, TTFT p50 3003.0448000179604 ms, E2E p50 3852.9262999945786 ms, and aggregate throughput 11.678387852350253 output tok/s. These measurements are short-output capacity evidence only. The current MTP text-only context scout completed 2/2 at tokenizer-observed inputs 32658 and 125838 with stop; native evidence is native/mtp-uva3-context-scout/artifact.json and its bounded summary is context-scout-summary.json. It is not the full matrix or a mixed text-plus-image context gate. The full 60-case text context matrix now passed 60/60 with all stop results at actual inputs 32658–32689 and 125838–125869, retaining a 4096-token output reserve. Synthetic agentic passed 2/2 across six API turns and four tool calls. This does not establish exact-limit/one-over admission, mixed text-plus-image context, C2, SWE, video, or soak behavior.
Native quality evidence deliberately retains source_recipe: null. Exact recipe-digest binding is supplied separately by identity-mtp-uva3.json, startup logs, and health evidence; do not invent fields in the native schema.
Before the first evaluation request, evaluation dispatch failed closed because the selected topology declared zero owners for the evaluation resource role. The retained failure is native/friction/evaluation-owner-failure.log; it did not reach the model endpoint. The durable correction is to invoke evaluation through a command-registry entry whose selected topology declares an evaluation owner and records its command host and runtime. Invariant: resolve and verify the resource owner before issuing an evaluation request; if no owner is declared, preserve the dispatch failure and do not attribute it to the candidate.
eval benchmark evidence show rejected a durable outer job artifact with schema benchmark-result/v1. Inspecting its unchanged nested results.evidence as benchmark-evidence/v1 passed with validation_errors=[]; both artifacts are retained. Invariant: retain the complete outer job artifact and inspect the nested canonical evidence rather than performing a lossy conversion. The related wrapper gap is recorded in product-gaps/2026-09-24-benchmark-evidence-job-wrapper.md.
Mixed context-plus-vision qualification remains a separate gate; no combined 128K image claim follows from distinct context and vision artifacts. The related gap is recorded in product-gaps/2026-09-24-mixed-context-vision-gate.md.
The no-MTP visual-only branch retained its own guarded page-cache reclaim, including the settled prior host sample at native/prior-host-settled/host-memory.json. It is historical and does not describe current MTP headroom.
The current MTP-UVA3 branch separately retained native/mtp-uva3/host-memory-after-reclaim.json after guarded idle reclaim. Invariant: reclaim only while guarded idle after model load, then verify delayed host recovery and endpoint health. A clean file-cache reclaim is never described as deleting model weights.
Managed CLI was used where MCP wrappers were unavailable. Preserve failures. The bounded direct 128K evidence does not promote a route, client alias, or deployment; exact-limit/one-over admission, mixed text-plus-image context, C2, SWE, video, and soak remain untested.