Skip to content

Model dossiers

Each dossier synthesizes immutable dated findings; it does not report live routes, placement, or availability. Start with the selection guide for a workload decision, or use this complete inventory to find the supporting evidence by measured hardware.

RTX PRO 6000

Model / configuration Measured result Limitation / decision Evidence date Dossier
Private-target cache-layout and DFlash2 K3 scouts 144/144 timed requests per trial; no 10% E2E win against both controls; cache slots +38.60% / −40.63% retain original; natural-output scout, private identity, no promotion 2026-11-01 Private runtime scouts
Nemotron 3.5 Super VL 120B-A12B NVFP4 TP2 scout preflight 6/6, vision 18/18, context 2/2; Oracle 48/54; C1 0/4 eligible exact Oracle and capacity prerequisite fail; no coding qualification, C4, speed, or promotion 2026-10-16 Nemotron 3.5 Super VL
2026-10-16 SGLang, Qwen, and Nemotron comparison terminal SGLang/Qwen SWE failures; bounded Nemotron functional evidence no candidate meets replacement gates; no performance comparison 2026-10-16 Campaign finding
Mistral Medium 3.5 TP2 bounded protocol and native reliability evidence r2 resource-policy and r3 campaign reliability rejection; no promotion 2026-10-15 Mistral Medium 3.5
DeepSeek V4.1 Flash Stage 2 compatibility-only, Stage 3 dummy-path, and Stage 4 bounded real-weight diagnostics full-role r4 failed its first Oracle8 round; no serving, quality, throughput, or promotion claim 2026-10-15 DeepSeek V4.1 Flash
2026-10-15 replacement evaluation completed replacement gates and bounded DeepSeek diagnostics no eligible challenger; exact GLM retained after restoration 2026-10-15 Campaign finding
GLM official-base TR3 EXL3 R11 retained Fresh baseline SWE 2/5; R19 bounded restoration accepted TP2/DCP1, 327K/C4, APC/no-spec/vision8; thinking-off gap; no challenger promotion 2026-10-13 GLM-5.3-Flash
MiMo V2.6 Flash RL TP2 no-spec preflight 25/25 and frozen SWE 4/5 Agentic profiles 15/18, 17/18, and 17/18 miss the 18/18 gate; no context, multimodal, C4, or speed measurement; not-qualified 2026-10-06 MiMo v2.6 Flash
2026-10-13 closure lanes GLM baseline 2/5; terminal Ignis 2/5 resolved/three empty; Qwen, DeepSeek, and MiMo retain bounded native outcomes No challenger clears complete 4/5 plus Django/Pylint; no promotion or cross-lane ranking 2026-10-13 Campaign finding
GLM-5.3-Flash TR3 4-bpw Abliterated TP2 preflight 25/25 and agentic 18/18 frozen SWE 3/5 below the 4/5 floor; no capacity or speed result; not-qualified 2026-10-06 GLM-5.3-Flash
GLM-5.3-Flash CSF/QAD TP2/DCP2 max functional 25/25; Oracle 18/18; initial SWE 4/5, then fresh-reload A1 2/5 attempts and 2/4 graded A1 misses the 4/5 floor; B1/B2/A2 stopped. Strict C4 0/4 is performance-ineligible; acceptance gates do not prove full-window, simultaneous-C4, or speed. rejected; exact r11 restored and readmitted 2026-10-05 GLM-5.3-Flash
Qwen prospective primary replacement frozen coding 3/5 strict C4 0/4 performance-eligible; evaluated profile rejected, no-promotion 2026-10-05 Qwen3.8 Flash Next
MiMo v2.6 Flash RL SGLang TP2 V3 healthy load, then two probes with zero visible content compatibility-only plus bounded negative correctness; no capacity/throughput or text/vision qualification; no-promotion 2026-09-21 MiMo v2.6 Flash
GLM-5.3-Flash mixed 3.5-bpw EXL3 Startup failed from host RAM exhaustion; no inference requests Rejected unbounded configuration; selected 4-bpw baseline restored; retry requires containment 2026-09-17 GLM-5.3-Flash
GLM-5.3-Flash mixed 3.5-bpw EXL3, R7 loader fix-forward C1/C4 functional, repeated C4 quality, 216K actual-token needle; one C4 4/4 cell 98.17 tok/s Strict capacity inconsistent; no overall win or promotion 2026-09-17 GLM-5.3-Flash
GLM-5.3-Flash EXL3 r7 no-spec 327,680 configured tokens; C4 strict120 120/120; separate long-context 9/9; MMLU-Pro 90/100, agentic 30/30, SWE 4/5 historical text-only quality lineage; no concurrent full-window soak or fresh-boot proof 2026-09-14 GLM-5.3-Flash
Qwen3.8 Flash Next EXL3 4.05-bpw MMLU-Pro 91/100, context 9/9, agentic 21/30, SWE 4/5, image/OCR 12/12 strict120 had zero performance-eligible responses; no-promotion 2026-09-14 Qwen3.8 Flash Next
Qwen3.8 Flash Next NVFP4 TP=2 253,703 actual prompt tokens plus 8,192 output; text, image, OCR, and video acceptance historical rollback evidence 2026-08-26 Qwen3.8 Flash Next
DeepSeek V4 Flash 0731 Retained TP=2 quality, capacity, and client evidence former Primary; historical 2026-08-21 DeepSeek V4 Flash
Qwen3.8 27B official FP8 Declared text-only rollback after the September selection unexercised after promotion; no fresh restoration proof 2026-09-14 Qwen3.8 27B
Qwen3.8 27B NVFP4 TP1 / TP2 / DP2 DP2 1,401.8–1,423.4 aggregate tok/s; TP1 764.3; TP2 587.9 TP2 strict JSON failures; all no-promotion 2026-09-04 Qwen3.8 27B
Agents-A1 Historical FP8 Primary-era evidence historical promotion 2026-07-29 Agents-A1
Qwen3.5 122B NVFP4 / MXFP4 Single-card qualification and TP=2 topology evidence retained rollback / no-promotion 2026-08-01 Qwen3.5 122B
Laguna S 2.1 and XS Qualified S and failed XS paths rollback / rejected 2026-08-01 Laguna S 2.1
GPT-OSS Puzzle 88B 40/40 c8 capacity, 20/20 tools, and long-context retrieval dated rollback; unified-diff only 2/3 2026-07-18 GPT-OSS Puzzle 88B
Gemma 4 official / Unsloth variants Official 12B passed strict quality and 240K retrieval; 31B 62.3 tok/s 26B timeout and 31B latency outcomes; no-promotion / rejected 2026-07-17 Gemma 4
Qwen3.6 27B and ThinkingCap Four-checkpoint comparison and quality control no-promotion 2026-07-13 Qwen3.6 27B
Nemotron 3 Super 120B Historical quality/capacity challenger no-promotion 2026-08-01 Nemotron 3 Super
Nemotron Puzzle 75B MTP 131K retrieval, 20/20 tools, and 137.0 tok/s MTP decode no-promotion; later planning recheck 0/5 2026-07-12 Nemotron Puzzle 75B
GPT-OSS 120B 128K retrieval, 20/20 tools, and 183.2 tok/s historical decode historical control; no-promotion 2026-07-12 GPT-OSS 120B
Mistral Small 4 131K context; 57.82 aggregate tok/s at c1 and 67.04 at c5 low-TTFT control; no-promotion 2026-07-12 Mistral Small 4
MiniMax M2.7 REAP 64K retrieval, 2/2 intelligence, and 97.2 tok/s no 131K headroom; no-promotion 2026-07-11 MiniMax M2.7 REAP
Ornith 1.0 35B 131K needle, 20/20 tools, and 29.2 tok/s incomplete identity; no-promotion 2026-07-10 Ornith 1.0 35B
Inkling Small NVFP4 Low-reasoning qualification lane no-promotion; Responses caveat 2026-08-01 Inkling Small

RTX 5090

Model / configuration Measured result Limitation / decision Evidence date Dossier
ThinkingCap Qwen3.8 27B AWQ 128K/C1; vision 18/18, limited quality 30/30, context 60/60, strict capacity 100/100 current after authorized secondary promotion; mixed near-limit vision and soak untested 2026-09-24 ThinkingCap Qwen3.8
FLUX.2 Klein 4B 6/8 bounded visual reviews pass workflow available=true, not promoted 2026-08-28 FLUX.2 Klein
Wan2.2 TI2V 5B Functional/decode and Hermes acceptance unavailable, no-promotion 2026-09-15 Wan2.2
Qwen3.8 27B GGUF / NVFP4 50.9 ms warm TTFT target-only and 137.7 tok/s 64K speculative arm challengers and incumbent retained; no-promotion 2026-09-03 Qwen3.8 27B
Gemma 4 E2B W4A16 30K/60K/120K preflight; 96 aggregate tok/s c1 and 204 c2 at 32K timeout triage 0/3; no-promotion 2026-07-16 Gemma 4 E2B
Gemma 4 31B W4A16 Fast lane 30K/45K/60K protocol preflight; 9/8 aggregate tok/s at 30K c1/c2 128K start lacked KV headroom; rejected; dossier also covers a separate PRO lane 2026-07-17 Gemma 4 variants
Signal 27B Q6_K/MTP3 Functional; 9/10 thinking-on quality scout 262K contract not qualified; strict output failures; no-promotion 2026-09-12 Qwen3.8 efficient variants
Swift 27B Q6_K/MTP3 Functional; 9/10 thinking-on quality scout 262K contract not qualified; strict output failures; no-promotion 2026-09-12 Qwen3.8 efficient variants
Qwopus Flash 27B Q6_K Functional; 1/10 thinking-on quality scout format and budget failures; no-promotion 2026-09-12 Qwen3.8 efficient variants
Minitron 20B Q6_K Functional; 3/10 thinking-on quality scout output and bounded-quality failures; no-promotion 2026-09-12 Qwen3.8 efficient variants
Qwen3.5 35B-A3B GGUF Q4_K_M 59,053 actual prompt tokens at the configured 64K window; 20/20 tools; 56.3 tok/s short probe historical fast candidate, no promotion and no immutable revision retained 2026-07-11 Qwen3.5 35B
Nemotron Nano/Omni 30B 20/20 tools, session recall, and 64K context on nightly lane historical topology record 2026-07-27 Nemotron Nano/Omni
Qwen2.5-Omni 3B 128K retrieval and 20/20 tools challenger, no-promotion 2026-07-27 Qwen2.5-Omni
Parakeet TDT 0.6B v3 Primary-human micro-WER 3.343% current in dated evidence 2026-07-28 Parakeet
Qwen3-ASR 0.6B Primary-human micro-WER 3.621% challenger, no-promotion 2026-07-28 Qwen3-ASR
Nemotron 3.5 ASR Primary-human micro-WER 6.685% rejected; 3.343-point WER regression 2026-07-28 Nemotron 3.5 ASR
Kokoro 289.27 ms TTS and RTF 0.1006 current in dated evidence 2026-07-28 Kokoro
Gemma 4 E4B Fast 30K retrieval, 20/20 tools, 61 ms warm TTFT, ~97 tok/s historical Fast control; no-promotion 2026-07-16 Gemma 4 E4B

Apple Silicon

2026-10-06 English voice campaign

Configuration Measured result Limitation / decision Dossier
Parakeet Metal control human WER 3.064%; warm C1 p95 81.68 ms critical entity failure; retained control, not newly qualified Parakeet
Qwen3-ASR 1.7B q8 human WER 2.786%; C1 p95 314.25 ms paired interval overlaps zero, critical error; no promotion Qwen MLX
Granite TurboCTC FP16 human WER 3.343%; C1 p95 40.16 ms critical/point regression; q8 static stop separate Granite
Kokoro 0.8.2 MPS warm 60/60; ASR proxy8.118% no blind naturalness/client qualification Kokoro
Breeze BF16 / q8 stream / q8 buffer 59/60resource abort;60/60;60/60 no human winner; exact configurations unqualified Breeze
Fish S2 Pro q8 buffered 56/60; four resource-aborted requests no partial ranking; streaming unsupported Fish
4B / stock9B / private 27B off/low / medium semantic 132/131/138/144/144 of 144 all strict 8K ineligible, exact failures/private fixture; no promotion Voice LLM MLX

Dated finding binds exact model revisions, runtime controls, all populations, failures and reconstruction limits. English Nemotron is a static integration stop, not a measured quality row.

Model / configuration Measured result Limitation / decision Evidence date Dossier
Qwen3.5-9B MLX 4-bit Functional 6/6; diagnostic quality 9/12 spoken 33/36, patch 0/3, output 0/10; no-promotion 2026-09-08 Voice LLM MLX
Qwen3.8-27B MLX 4-bit Compatibility preflight 5/6 tools 1/3; incomplete, no-promotion 2026-09-08 Voice LLM MLX
Qwen3.6-35B-A3B MLX 4-bit Spoken 36/36 and one SDK session strict JSON failure; incomplete, no-promotion 2026-09-08 Voice LLM MLX
Qwen3 4B MLX 4-bit baseline Spoken suite 36/36 strict shared-prefix tools 0/3; unchanged baseline, not a promotion 2026-09-08 Voice LLM MLX

The result and decision labels are historical evidence labels. Failed loads and compatibility-only runs remain visible; none of these rows authorizes serving or routing changes.