Configuration and identity¶
All measured lanes used RTX PRO 6000 Blackwell Max-Q hardware. TP1 uses one GPU; TP2 uses two. The machine-readable bindings retain exact public model revisions and managed serving-image pins with source hashes. They are a derived allowlist, not a native log or new run. Native evidence links below bind the observed requests. SWE paired-image IDs identify grader containers and must never be used as serving-image identities.
Exact model and runtime pins¶
| Lane | Model repository and revision | Serving runtime / immutable image |
|---|---|---|
| GLM retained R11 | brandonmusic/GLM-5.3-Flash-tr3-4bpw, a5fee929cf4888b1824323e33e8a19b60129e025 |
Custom vLLM 6dc2f516688fe6f84c6994dcd20fddf296853a6c; image ID sha256:0f1cdcc8891f1cc3a444121eb61d366289a1cbba285f0892dcbb24bc94961692 |
| Coder390 INT4 / INT8 | nerkyor Coder390, ccec60c489c1d007704d38de436ec7019457edf8 |
vLLM 0.28.0; vllm/vllm-openai:v0.28.0@sha256:2286e8533ca8b6bc777594bae30524f1426ba46ca21797524e06df6a94b06635 |
| Mixed GLM | ormandj/GLM-5.3-Flash-W4A16-NVFP4-K32-Experts-FP8-WO, c3cbb9891b67c741bcbf6b176dd7af9265b069db |
Custom SGLang runtime release 0.4.3 rc.6, SGLang source 00143e9c23aee2dead5e6fe217bda4fa8739cb92; ghcr.io/ormandj/sglang-glm53-flash-sm120@sha256:ec4243f940a179a27fea21895077efd47cd050501f99a1d2a5fecf7df2e7be71 |
| Qwen QAD | local-inference-lab/Qwen3.8-Flash-Next-NVFP4, 6909a5bed089a48fa07e956d3915af2537de9368 |
LIL vLLM Karmic Kraken; managed pin ghcr.io/local-inference-lab/vllm:karmic-kraken-beta-20261005-e1ccbaf0fc0f4f4c@sha256:6008f020c23065bef71390194f968af9f5ebbedf57807b1659f876632e4a1067; TP1 recipe records engine source 00a33e23142097f99fef95b1ecab6da4ea271e2e |
| DeepSeek EXL3 | turboderp/DeepSeek-V4-Flash-0731-exl3, b5526babe907141a0750f265a5a2a5cf414f5a64 |
TabbyAPI/ExLlamaV3; Tabby source a3fed43efae3b18ec6181618d6a4a66153a9d14e; ghcr.io/theroyallab/tabbyapi:cu13@sha256:bafb7fae72ce586d6292db1947b7190b2165c43aed2bf15f879cf603637ed826 |
| Qwen EXL3 schema correction | turboderp/Qwen3.8-Flash-Next-exl3, 55a732e0c4c3d4614bc42b68493bb930d9b02c0a |
Same Tabby base source plus schema-aware operator patch 997fda896d0f994eee3b97ea212d850fbbbd76f58c2bfbdbc667f3ac30a09502; derived image ID sha256:b61766e6c60749d0695bce6cdee1a4ceb5886ff9dbd45e5667243cf4c24447c7 |
| Stock Qwen27 FP8 | Qwen/Qwen3.8-27B-FP8, 017b9c7af6b5689d5dd426a76e0bc077eb5ca20a |
SGLang 0.5.19 0bcd822377da7b5718e674eaf9c870d349424dd1; lmsysorg/sglang@sha256:37bbbd3444732a464bbc68dee4fb0164e0ce9e18e2f027f3fc967f1152d3c262 |
| MiMo MOPD | ProCreations/MiMo-V2.6-Flash-MOPD-NVFP4, da654d8ed8d0d2acbf0fc5b63df72a7b75a6bb2a |
vLLM 0.28.0, same exact 2286e853… image above; six load trials, no inference queries |
| Ignis R4 | gpillon/Qwen3.8-Flash-Next-trellis-a25-ignis, f75e492c6d726537a4d4c85858bc91ee06a070ef |
Ignis 0.6.2; ghcr.io/gpillon/ignis@sha256:3144a2ed2240969343e2378c518232520a994833f364235b789a3e8d4b6bcdba |
The GLM lineage receipt identifies the official BF16 teacher zai-org/GLM-5.3-Flash-BF16@a6c167b62691b2bac901344b65cb651a70f53e43. The deployed/restored weights are TR3 EXL3 4 bpw, not BF16 and not the abliterated derivative. This is publisher provenance plus retained receipts, not a fresh teacher-to-quant tensor audit. Runtime labels are retained in the bindings as labels; an inherited label alone does not attest patched runtime contents. The exact bundled ExLlamaV3 core revision remains unverified where stated by native controls.
Runtime envelope and actual request controls¶
Configured context is not measured prompt length. Sampling below describes SWE/quality requests; protocol preflight defaults differ and are retained separately.
| Lane | Weight / cache / placement | Configured TP, context, concurrency | Actual request controls and evidence |
|---|---|---|---|
| GLM R11 | EXL3 4 bpw, FP8 DS-MLA KV; APC, no speculation | TP2/EP2/DCP1; 327680; C4; vision 8 | SWE T0.6/p0.95, thinking enabled; native SWE. No new sampler/kernel optimization. |
| Coder390 INT4 | INT4/W4A16 subtree; GPTQ W4A16, symmetric INT4/group128; BF16 retained components; KV auto, 16 GiB configured pool; no MTP |
TP1; 131072; C1 | Initial thinking enabled; completed medium effort T0.6/p0.95; native medium. TP1 is authoritative despite old preparation filenames. |
| Coder390 INT8 | INT8/W8A8 subtree; SmoothQuant W8A8, per-channel INT8 weights/dynamic per-token INT8 activations, native Triton; KV auto; no MTP |
TP1; 131072; C1 | Medium T0.6/p0.95; native control. |
| Mixed GLM | Mixed NVFP4/K32 experts with FP8 WO; FP8 E4M3 KV, BF16 Mamba state; no speculation | TP2; 131072; C1 | Separate enabled/off arms, both T0.6/p0.95; enabled SWE, off SWE. Prepared 327K/MTP recipes are not completed results. |
| Qwen QAD | NVFP4 routed experts/PLE, MXFP8 attention/shared experts; FP8 KV, FP32 recurrent state; full PLE; no MTP | TP2; 262144; C1. Contained TP1 131072 is compatibility-only | Xhigh T1/p0.95; TP2 SWE. TP1 supplies no SWE or performance result. |
| DeepSeek EXL3 | EXL3 2.04 bpw; FP16 KV; no MTP or CPU MoE offload | TP1; 131072; C1. TP2 32768 compatibility-only | Low/max T1/p0.95, native top-k 0; native max binding. |
| Qwen EXL3 | EXL3 4.05bpw_h6_ng6; FP16 KV; streamed PLE requested; no MTP | TP1; 131072; C1 | Xhigh/medium T1/p0.95, actual native top-k 0; xhigh binding, medium binding. Top-k 20 arm cancelled before load/query. Full-input/fully-GPU PLE unverified. |
| Stock Qwen27 | Official FP8 weights; corrected BF16 KV, FP32 Mamba reduction; no speculation | TP1; 131072; C1. TP2 compatibility-only | Xhigh T1/p0.95; inherited top-k20; native binding. Initial FP8-KV lane is a distinct diagnostic, not a matched precision comparison. |
| MiMo MOPD | NVFP4; text-only trials with zero/2 GiB selective offload; KV 1.5 GiB planned | TP2; 32768; C1 planned | Six failed loads, no request sampler/quality result. Required 4 GiB reserve retained; TP1 95.9 GiB CPU case pruned. Final failure decision. |
| Ignis R4 | Trellis mixed 2–4 bpw, mean 2.5; roughly 40 GB streamed CPU experts, not GPU-only; cache dtype not independently measured | TP1; 131072; C1 | SWE T1/p0.95/default thinking; explicit sampler control top-k20 is separately retained. Actual managed load. |
Preflight erratum: the native source-062 protocol probes used T0 and omitted request top-p/top-k; earlier control prose saying native T1/p1 is inaccurate. Streaming probes also omit flat effort and can inherit native xhigh. Do not infer every protocol wire used an explicitly requested effort, or copy protocol defaults into SWE/quality rows. The bindings preserve the thirteen actual SWE request-control objects. Ignis C8 preflight (6/13), serial C1 controls (6/6), quality (12/12), and recovered SWE retry notices (24) are separate populations.
R19 restoration accepted the exact retained baseline and bounded client checks. No challenger was promoted; operator topology, endpoints, host names, GPU UUIDs, and runtime container IDs remain private.