One real production workload. Twenty-five instrumented runs across two open-weights families and four generations of Claude. Mining GPUs, thermal forensics, and receipts for every number. These are the notes, posted raw.
Driving Chrome in an agentic loop, Qwen 3.8 Flash-Next navigates to a signed Alibaba Cloud object-storage URL it invented: 235 times in 17,679 navigations. Under identical workloads and prompts, Claude, GPT-5.6 Sol, GLM-5.3 and DeepSeek do not. It also reports sending such URLs when it did not.
Three runs of a self-hosted INT8 checkpoint on a patched engine average 77.0 — above the rented llama.cpp anchors and above Claude Sonnet 5, at about a dollar of electricity a run. The same weights scored 64.5 before the sampler was fixed.
vLLM's Triton top-k/top-p kernel drops the weakest surviving candidate when min_p, top_k and top_p are used together. Nine days of runs decoded near-greedily, every correctness instrument passed, and the investigation started on the weights.
A tensor-by-tensor comparison of the Unsloth Q8_0 GGUF and the official FP8 checkpoint against the BF16 weights, three purpose-built INT8 checkpoints to close the gap, and a corpus measurement that cannot resolve any of them.
The same model scored 11 points lower on vLLM's Anthropic bridge than on llama.cpp. The bridge hands the model its own prior reasoning back, it imitates itself, and within a few hundred steps 40% of its thinking is one 62-character sentence.
Nine model configurations — two self-hosted open-weights, four Claude generations, GPT-5.6 Sol, DeepSeek and GLM — each ran the same long-horizon enrichment pipeline three times, scored against adjudicated ground truth with weights fixed before results.
The chassis RGB lives in SMM firmware with no Linux driver in existence. A write designed to change nothing turned the lights off — and seven confident beliefs died in 27 hours.
The capstone: the full pipeline on a 4090+3090 pair — faster than two of three rented datacenter runs, for 68 cents of electricity. Plus the +17% environment variable and the wall that stopped a bigger model.
A 3.5-slot GPU leaves 5 mm above the next slot. Every riser on the market needs 8–15. A tour of CAD teardowns, a vendor with no drawings, and the fix that made it all moot.
ATX 3.1 quietly gutted PCIe 8-pin counts — verified unit by unit. Plus the mining-card power socket that looks exactly like the connector that would destroy it.
My 3090 sits 3 mm above a 4090. Under sustained load its core read a comfortable 70°C — while the memory junction, invisible to standard tooling, ran 94. Getting that number took reverse-engineering a monitoring app's shared memory.
Two configurations swapped rank when third replicas landed — one of them spanning 15.8 points between identical runs. Single-run model comparisons measure the dice.
Transcript autopsy of every run that stopped without finishing: a model that mimicked its own memory system's voice, a promised file that never appeared, and the harness semantics that let both count as 'done.'
85% of my agents' tokens were spent rebuilding memory they'd been forced to throw away. One environment variable and a native context window made runs 7× faster — and better.
A third of Qwen's flash model is a predictive n-gram table that never does a matmul. It quantizes as a step function, offloads to RAM for free, and even runs from SSD.
The CMP 170HX is A100-class silicon that unlocks to 64 GB of HBM2e. Buying two meant mapping a market that is part bargain, part minefield.
Identical weights ranked last in one agent stack and first in another. Most leaderboard deltas I measured live in the plumbing, not the parameters.