# TensorQuay lab-note data Reviewed 21 September 2026. These are selected numeric measurements from our historical tests, with units and settings. They are not a general hardware rating or a standard model-quality benchmark. ## FastH3 V2 / one RTX 5090 Article: https://tensorquay.com/rtx-5090-video-generation.html `fasth3-timings.csv`: 18 September 2026, FastH3 V2 pruned INT8 ConvRot, eight steps, VSA keep 10% starting at 20% of the schedule. One RTX 5090 with 32 GB VRAM and 125 GB host RAM. Four fixed prompts, one measured run per prompt, per warm cell. Warm-ups excluded. `observed_wall_seconds` lists all observations in that cell so its median can be recalculated. Frames are at 24 fps. Time runs from prompt submission to saved MP4; model download is excluded. Cold start is one separate observation after unloading the models, not a cold-start median. Read the article for text encoder, decoder, memory and quality limitations. ## GLM-5.3-Flash / two RTX PRO 6000 GPUs Article: https://tensorquay.com/glm-5-3-flash-dual-rtx-pro-6000.html 19 September 2026. Two 96 GB RTX PRO 6000 Blackwell Server Edition GPUs (600 W variant), 503 GB host RAM. Unsloth UD-IQ4_XS on llama.cpp `86ebfef`, layer split, microbatch 512, thinking on, temperature 1.0, top-p 0.95. - `glm-scripted.csv`: one batch per concurrency/MTP setting, short prompts and 512 output tokens per request. Decode speed is the median server-reported per-request speed. Aggregate speed uses the whole batch wall time. First-token time is measured at the client. These are not long-context coding-task results. - `glm-agents.csv`: one wave per setting, small Python/Rust/HTML tasks, basic file checks, external MCP integrations disabled. `finished_normally` is distinct from `task_checks_passed`; capped/error runs can leave passing files. The 4/8/16-agent waves use 16 slots × 64K; larger waves use the context stated in the CSV. Recorded time includes request waits, tools and harness polling/check overhead. Medians include every started task, including capped/error runs; they are not medians of successful completions alone. - `glm-agent-observations.csv`: individual elapsed times and fixed outcome labels, sufficient to recalculate the medians and counts. Observation numbers only distinguish rows within a wave. - `glm-iq4-sha256.txt`: the five model-shard hashes verified for this experiment. The website does not distribute the weights. A download from the model repository's current revision must match all five hashes to reproduce these inputs. For an even-sized group the median is the average of its two central sorted values. The original agent harness chose the upper middle value. The article and these CSVs correct that summary error using the individual records: MTP-off medians are 158.75, 272.0, 702.8, 1119.85 and 1775.35 seconds for 4/8/16/32/64 agents. The 16-agent MTP-on median is 716.45 seconds. A task-check pass is not proof of general model quality or quantisation parity. ## Qwen3.8-27B / one RTX 5090 Article: https://tensorquay.com/rtx-5090-local-ai-benchmark.html 2 September 2026. Qwen3.8-27B NVFP4, vLLM 0.28, FP8 KV, MTP off. - `qwen-cache.csv` transcribes matched configurations from retained experiment 8a/8c tables. GPU-only and 24 GiB CPU-tier configurations each have two successive warm-continuation repeats. Histories are the recorded 8K/16K/32K labels; continuation adds 12 input tokens and generates 128 output tokens. Effective per-session speed includes request wall time. These are summary-table values, not a newly reprocessed per-request dataset or statistical confidence bounds. The cache experiment used a 116 GiB container-memory cap. - `qwen-agent-times.csv` transcribes the per-session completion times from experiments 9h/9j, sorted by time and without task text. Sixteen tasks completed in the 16-session wave; 27 completed in the 30-session wave, with two harness failures and one unfinished task excluded from the completion-time median. The standard medians are 447.5 seconds (7.5 minutes) and 1240 seconds (20.7 minutes). The waves mixed continuing and fresh tasks, so their timing difference does not isolate concurrency alone. Completion is not an independent quality-check pass. ## Qwen3.8-27B / long sessions / one RTX 5090 with a 64 GB-class RAM budget Article: https://tensorquay.com/qwen-long-sessions-rtx-5090-64gb.html `qwen-long-sessions-64gb.csv`: our in-house RTX 5090 (32 GB) test machine with a 64 GB-class RAM budget (62 GB usable); not an assembled TQ–32. Runs `A1` (26 September 2026) and `A2` (28 September 2026): Qwen3.8-27B NVFP4, vLLM 0.28.0, FP8 KV, MTP off, 24 GiB CPU KV-cache tier (`--kv-offloading-size 24`). Each session's history is filled one at a time; then all sessions send a short follow-up together and generate 128 tokens, twice (`repeat` 1 and 2). History is synthetic filler text. Tok/s per user is 128 divided by that user's wall-clock time, including queueing, cache restore, prefill and decode. `per_user_tps` lists every user so the mean, median and range can be recalculated. Run `NVMe` (28 September 2026, one run) used vLLM 0.30.0 with the same 24 GiB RAM tier plus a second tier on the machine's NVMe SSD, 12 sessions with unique random-token 64K histories; it is not directly comparable with runs A1/A2. No run measures agent task success. Neither the Qwen nor FastH3 test certifies the thermals, noise, power use or workload capacity of an assembled TQ–32. GLM used a different two-card server. Further tests need repeated tasks and matched conditions before making broader comparisons.