Skip to content

Xeon model qualification campaign

Field Result
Date 2026-10-05
Scope prospective Qwen replacement and GLM CSF/QAD compatibility scout
Measured hardware two RTX PRO 6000 Blackwell Max-Q cards
Baseline / frozen replacement floor selected GLM r11 DCP1; 4/5 SWE
Candidate outcome Qwen 3/5 rejected; GLM initial max-only SWE 4/5, but fresh-reload A1 repeat 2/5 attempts (2/4 graded)
Functional headlines Qwen medium 25/25; GLM max functional 25/25; GLM max-only Oracle 18/18
Strict C4 headline baseline, Qwen, and GLM each 0/4 performance-eligible
Evidence sanitized artifact bundle

This campaign rejects both evaluated replacements and retains selected GLM r11 DCP1. Managed restoration completed with direct 25/25, routed 7/7, and a fresh Pi read-tool turn returning an unseen fixture value; routes and manifest hashes were unchanged. It does not establish a hardware speed result, long-context capacity, or a qualified replacement.

Public reconstruction

All probes used a sanitized direct-endpoint reference. The published evidence retains only its endpoint class and served-model identity, not an address. “Configured” context and C4 are setup values, not measured long-context or concurrent-capacity results.

Cell Engine / quant / KV TP / DCP / APC / speculation Context / output / images / max batched tokens Containment and reasoning
Baseline baseline-local-tier-r11 pinned GLM EXL3 4-bpw / FP8 DS-MLA KV TP2 / DCP1 / APC enabled / no speculation C4; 327,680 context; 65,536 configured output; 8 images, 0 video; max batched tokens 2,048 52 GiB RAM + 1 MiB swap; request controls are retained separately in each native artifact
Qwen candidate-local-tier-r1 vLLM 00a33e… / ModelOpt mixed / FP8 KV TP2 / DCP1 / APC enabled in recipe / no speculation C4; 262,144 context; 65,536 configured output; 8 images, 0 video; max batched tokens 6,019 52 GiB RAM + 1 MiB swap; 51.58 GiB/rank loaded; medium functional profile
GLM glm53-csf-qad-tp2-dcp2-c4-327k-r1 vLLM 00a33e… / NVFP4-CSF with MXFP8 attention/shared experts / FP8 KV TP2 / DCP2 / APC enabled in recipe / no speculation C4; 327,680 configured; recipe output cap 65,536; 8 images, 0 video; max batched tokens 4,096 52 GiB memory + 1 MiB swap; default HIGH, corrected Oracle/SWE uses thinking_mode=default + reasoning_effort=max

The configuration record retains all three exact model revisions, image digests, and configured limits. Completion caps used by individual probes can be smaller than the configured output ceiling. These values do not establish a measured full-window envelope.

The baseline reconstruction is from the selected r11 DCP1 finding; the Qwen and GLM rows are from their linked startup identities. The common runtime image and the candidate revisions are pinned in the source registry.

The research dispositions retain candidate leads and exclusions. Artificial Analysis values are hosted/API priors, not local quantization or hardware evidence. The pinned runtime's host-RAM cache documentation conflicts with launcher behavior that limits recurrent restore to text; VRAM-only cache does not resolve that multimodal boundary.

Retained results

The evaluated Qwen replacement profile, local-inference-lab/Qwen3.8-Flash-Next-NVFP4@6909a5bed089a48fa07e956d3915af2537de9368, passed disabled preflight 2/2, medium preflight 25/25, and Oracle 18/18, then was rejected at 3/5 against the frozen r11 4/5 primary coding floor. Its strict C4 128-word cell completed 0/4 performance-eligible requests. Those results reject that evaluated profile; they do not make a family-wide claim about Qwen.

The GLM candidate was local-inference-lab/GLM-5.3-Flash-NVFP4-MXFP8-CSF-QAD at revision 4f90b74165cb1542cca35730b7dfa4d550bcf0ce, using the pinned runtime image sha256:6008f020c23065bef71390194f968af9f5ebbedf57807b1659f876632e4a1067. The startup record shows TP2/DCP2, C4, 327,680 configured tokens, disabled speculation, and no OOM event. It is compatibility evidence only; 327,680 and C4 are configured values, not measured envelope claims.

GLM disabled-thinking preflight passed 2/2 and the pinned HIGH profile preflight passed 25/25. The first Oracle run passed 17/18; its only failure returned a Markdown fence where structured-edit required JSON. The strict C4 128-word cell completed 0/4 performance-eligible requests, so it supplies no throughput or speed result.

The original explicit enabled-thinking plus max request failed before inference because those controls conflict. A corrected, separate Oracle run uses thinking_mode=default with reasoning_effort=max, which derives the pinned API's effective enabled-thinking behavior and passed 18/18. The same max controls passed functional preflight 25/25 and the single five-task SWE scout resolved 4/5; only Django failed. That meets r11's absolute replacement floor, but differing reasoning controls prevent a quality-improvement or causal comparison claim.

The candidate passed four exact eight-image requests. It also passed 4/4 context probes at 247,617–247,638 actual input tokens with a 65,536-token maximum, and 4/4 near-full probes at 316,766–316,774 actual input tokens with an 8,192-token maximum. These are acceptance results, not generated-output lengths or a full 327,680-token request. An offered-C4 long summary passed 4/4 with own canaries, no foreign output, and terminal completion at 165,903–165,908 actual input tokens (nominal 201,000). It does not establish simultaneous occupancy or speed.

Independent review locked a fresh-reload ABBA comparison: candidate A1, incumbent B1/B2, candidate A2. Every period required Oracle 18/18 and SWE at least 4/5; its sole comparative metric would have been SWE resolved tasks per agent-stage hour, excluding grader and preparation time, with paired and aggregate improvement of at least 10%. A1 failed the quality stop: of five attempts, four were graded and only Sphinx and Xarray resolved; Django and Pylint graded false, and Sympy reached the 60-step limit ungraded. The native graded resolve rate is 2/4 = 0.5, and the campaign result is 2/5 attempts, below the 4/5 floor. The earlier 4/5 scout remains retained but does not show a reliable floor. B1, B2, and A2 did not run, so no comparative speed result exists.

During matched candidate A1 SWE work, the upgraded host exposed PCIe Gen5 x16 on both cards under actual model load (100% GPU utilization, 292.24/286.35 W under 300 W limits, and 72/74 °C). This point-in-time observation is not PCIe bandwidth, comparative speed, or thermal-peak evidence.

Closure status

  • Exact r11 restoration is complete: its pinned model, image, recipe, and registry were restored; direct 25/25, routed 7/7, and a fresh Pi anvil/llm.primary read-tool turn passed. Both router tiers were ready and admitting, their router/model-manifest hashes were unchanged, and only restored-model GPU workers remained.
  • Independent review now reviews the retained failure and closure evidence; it does not reopen candidate advancement.
  • Strict-C4 remediation, the explicit served-name recipe flag, and the host-RAM-cache trial are unexecuted leads. They are not authorized follow-on changes from this campaign.
  • This finding authorizes no model, route, lifecycle, or hardware change.

Evidence