https://agmind.ru/ Qualification research on local AI systems: evidence, limitations and negative results. en-US https://agmind.ru/essays/benchmark-that-sounds-like-a-human/ https://agmind.ru/essays/benchmark-that-sounds-like-a-human/ Our first chat corpus asked the model about build digests and goodput. Nobody's users ask that. The same cell re-measured on sixteen everyday prompts gave a different failure rate on identical hardware, and we kept the old number anyway. Wed, 02 Sep 2026 00:00:00 GMT https://agmind.ru/essays/ci-fails-when-a-number-drifts/ https://agmind.ru/essays/ci-fails-when-a-number-drifts/ Type a tokens-per-second figure into this site and CI fails. Every measured value is re-derived from raw run records, and a committed value that drifts from them stops the pipeline. How the gate works, what it enforces, and the class of error it cannot see. Wed, 02 Sep 2026 00:00:00 GMT https://agmind.ru/essays/independent-proof-or-it-didnt-happen/ https://agmind.ru/essays/independent-proof-or-it-didnt-happen/ DeepSeek opened its agent harness and named the mode, effort and sampling behind V4-Pro's agent scores. Better disclosure than the norm, and still not proof: a vendor's number stays vendor-disclosed until a second party runs the same config and publishes the traces. Wed, 02 Sep 2026 00:00:00 GMT https://agmind.ru/essays/agentic-coding-prefill-problem/ https://agmind.ru/essays/agentic-coding-prefill-problem/ Ask how many tokens per second a local coding agent needs and every answer quotes decode speed. But an agent's turn is a huge prompt and a short reply — the wait lives in prefill, and the hardware conversation is optimizing the wrong half of the request. Tue, 01 Sep 2026 00:00:00 GMT https://agmind.ru/essays/agents-per-box/ https://agmind.ru/essays/agents-per-box/ The agent-swarm conversation runs on single-stream decode numbers, which is the one metric that stops mattering the moment you run a swarm. What we learned sizing shared boxes, and the question buyers should be asking vendors instead. Tue, 01 Sep 2026 00:00:00 GMT https://agmind.ru/essays/harness-changelog-of-distrust/ https://agmind.ru/essays/harness-changelog-of-distrust/ A benchmark harness earns trust by what it declines to measure. Every version of ours is a scar with a commit message: the invalid first run we kept, the neighbor check, the fallback detector, the word that matched inside another word. Tue, 01 Sep 2026 00:00:00 GMT https://agmind.ru/reports/beelink-gtr9-pro-local-llm/ https://agmind.ru/reports/beelink-gtr9-pro-local-llm/ Reviews of this box benchmark it for a week. We bought two, froze the workloads, and have been serving from them since: the software stack that works, what the numbers look like, how the two units compare, and the parts we still cannot vouch for. Fri, 21 Aug 2026 00:00:00 GMT lab_repeated https://agmind.ru/reports/dgx-spark-256k-vllm/ https://agmind.ru/reports/dgx-spark-256k-vllm/ The 256K-context configuration hunt on a single GB10: which vLLM attention and MoE backends survived SM_121, why NVFP4 decoded slower than FP8 in the mainline build we pinned, the patched setup that held 260K — and the measured numbers, dated to the versions named inside. Fri, 21 Aug 2026 00:00:00 GMT lab_single_run https://agmind.ru/reports/local-llm-homelab-hardware/ https://agmind.ru/reports/local-llm-homelab-hardware/ Every hardware guide for local AI is a spec sheet, a vendor reviewing itself, or an affiliate listicle recycling VRAM tiers. This one starts from the workload, uses only numbers we measured on hardware we own, and marks unmeasured lanes as gaps. Fri, 21 Aug 2026 00:00:00 GMT lab_repeated https://agmind.ru/reports/russian-embedder-24m-params/ https://agmind.ru/reports/russian-embedder-24m-params/ The story behind strizh-ru-retriever: layer-pruning RuModernBERT-small into a four-layer encoder, the distillation shortcut that produced a beautiful loss and dead retrieval, what the compression kept — and the leftover config value that silently truncated every input. Fri, 21 Aug 2026 00:00:00 GMT lab_single_run https://agmind.ru/reports/russian-rag-splitter/ https://agmind.ru/reports/russian-rag-splitter/ The story behind agmind-rag-splitter-ru: why a chunker that rewrites text is dangerous for RAG, how a LoRA fine-tune learned to return boundary indices as JSON, the training corpus published alongside it — and the honest gap between teacher agreement and retrieval quality. Fri, 21 Aug 2026 00:00:00 GMT lab_single_run https://agmind.ru/reports/strix-halo-memory-allocation/ https://agmind.ru/reports/strix-halo-memory-allocation/ The classic first failure on a Ryzen AI Max+ 395 box: the model does not fit, or the runtime reports a few gigabytes of VRAM on a 128 GB machine. The BIOS split is usually the wrong lever. Here is what our serving node actually reports, and the setting that matters. Fri, 21 Aug 2026 00:00:00 GMT lab_single_run https://agmind.ru/reports/two-node-home-ai-lab/ https://agmind.ru/reports/two-node-home-ai-lab/ Our two identical Strix Halo units live opposite lives: one carries the whole containerized service stack, the other serves a single model and stays otherwise empty. The split is not tidiness — the benchmark harness invalidates any run that finds a GPU neighbor on the host. Fri, 21 Aug 2026 00:00:00 GMT lab_single_run https://agmind.ru/reports/which-inference-engine-local-llm/ https://agmind.ru/reports/which-inference-engine-local-llm/ The advice circulating is written from datacenter GPU benchmarks and rarely survives contact with a single box. What each engine is actually for, which one we run where and why, and the defaults question that matters more than the choice itself. Fri, 21 Aug 2026 00:00:00 GMT lab_single_run https://agmind.ru/reports/do-local-llms-invent-answers/ https://agmind.ru/reports/do-local-llms-invent-answers/ Everyone asks whether a local model hallucinates over their documents; almost nobody measures it. Our workload has an unanswerable control built in: questions the document cannot answer, graded on whether the model admits it. Two model families, two ways to not know. Thu, 20 Aug 2026 00:00:00 GMT lab_repeated https://agmind.ru/reports/how-many-tokens-per-second-is-enough/ https://agmind.ru/reports/how-many-tokens-per-second-is-enough/ The most asked and least measured question in local AI, usually answered with invented round numbers. What speed actually feels like on a real box: the wait before the first useful word, the flow after it, and the failure that makes tokens per second irrelevant. Thu, 20 Aug 2026 00:00:00 GMT lab_repeated https://agmind.ru/reports/how-to-benchmark-local-llm/ https://agmind.ru/reports/how-to-benchmark-local-llm/ Most published local-LLM speeds cannot even be disagreed with: they carry no configuration. A practical protocol from a lab that publishes raw runs — what to pin, what to ask the running server, which tokens to count, and why failures belong in the denominator. Thu, 20 Aug 2026 00:00:00 GMT lab_single_run https://agmind.ru/reports/local-llm-json-reliability/ https://agmind.ru/reports/local-llm-json-reliability/ Strict-JSON output is the load-bearing wall of local automation, and its reliability is usually asserted, not measured. On one box: quantization did not move it, the backend did not move it, reasoning mode did not move it — switching the model family did. Thu, 20 Aug 2026 00:00:00 GMT lab_repeated https://agmind.ru/reports/should-you-disable-thinking-local-llm/ https://agmind.ru/reports/should-you-disable-thinking-local-llm/ Reasoning mode on a local model delays the first visible word from a blink to a wait and multiplies the cost of automation — and on every task with a hard right answer in our gates it bought nothing. The measured trade, and where it does not apply. Thu, 20 Aug 2026 00:00:00 GMT lab_repeated https://agmind.ru/reports/strix-halo-vs-dgx-spark/ https://agmind.ru/reports/strix-halo-vs-dgx-spark/ The most searched comparison in the local-AI mini-box class is answered everywhere by spec sheets and nowhere by one lab measuring both. We own both platforms. Here is what each one demonstrably did, which workload belongs on which box, and the comparison we refuse to fake. Thu, 20 Aug 2026 00:00:00 GMT lab_repeated https://agmind.ru/reports/what-128gb-unified-memory-runs/ https://agmind.ru/reports/what-128gb-unified-memory-runs/ The 128 GB unified-memory class is the current sweet spot for local LLM boxes, and almost everything written about it is spec-sheet arithmetic. Here is what one such box measurably does: which model class it serves, how fast, for how many people, and where 128 GB ends. Thu, 20 Aug 2026 00:00:00 GMT lab_repeated https://agmind.ru/essays/benchmark-number-config-in-disguise/ https://agmind.ru/essays/benchmark-number-config-in-disguise/ Two honest people measure the same model on the same hardware and disagree hard. Nobody lied: a published speed is a compressed configuration, and the compression is lossy. What a number must carry to be worth quoting, and how to read one that carries nothing. Thu, 20 Aug 2026 00:00:00 GMT https://agmind.ru/essays/dashboard-lies-both-directions/ https://agmind.ru/essays/dashboard-lies-both-directions/ One box, one afternoon: a model returning empty answers under HTTP 200, and two healthy containers marked unhealthy by a probe aimed at the wrong port. Green-but-broken and red-but-fine are the same disease, and the cure is the same too. Thu, 20 Aug 2026 00:00:00 GMT https://agmind.ru/reports/deepseek-v4-flash-0731-dgx-spark-recipe/ https://agmind.ru/reports/deepseek-v4-flash-0731-dgx-spark-recipe/ The exact serving recipe behind our published GB10 numbers: weights and image pinned, the head-first launch order that avoids a TCPStore hang, the MoE backend flag auto-selection skips, what to monitor when NVML goes silent — and the numbers to expect. Sun, 16 Aug 2026 00:00:00 GMT lab_single_run https://agmind.ru/reports/dgx-spark-gpu-monitoring/ https://agmind.ru/reports/dgx-spark-gpu-monitoring/ The standard GPU observability stack half-works on GB10 unified memory: dcgm-exporter does not run and NVML returns N/A on a large share of queries, so Grafana stays green and empty. The collector we actually serve vLLM with, and how to build it. Sun, 16 Aug 2026 00:00:00 GMT lab_single_run https://agmind.ru/reports/dgx-spark-worth-it/ https://agmind.ru/reports/dgx-spark-worth-it/ The highest-volume buyer question in this niche is answered almost entirely by spec-sheet arithmetic. Here is what a pair of GB10 nodes actually did in our lab — serving a 284B MoE, walking a context curve to a million tokens — and who should and should not buy one. Sun, 16 Aug 2026 00:00:00 GMT lab_single_run https://agmind.ru/reports/llamacpp-prompt-cache-strix-halo/ https://agmind.ru/reports/llamacpp-prompt-cache-strix-halo/ A document Q&A session on a Ryzen AI Max+ 395 box: what the first question over a 32k-token file costs, what the second one costs with the prompt cache on and off, and the one condition the cache needs before it saves you anything. Sun, 16 Aug 2026 00:00:00 GMT lab_repeated https://agmind.ru/reports/q8-0-vs-q4-k-m-strix-halo/ https://agmind.ru/reports/q8-0-vs-q4-k-m-strix-halo/ Same model, same llama.cpp build, same Strix Halo box, only the quantization changed: what Q8_0 costs in decode pace against Q4_K_M, what our quality gates saw, and which shortage should actually drive the choice. Sun, 16 Aug 2026 00:00:00 GMT lab_repeated https://agmind.ru/reports/strix-halo-sustained-load-3-hours/ https://agmind.ru/reports/strix-halo-sustained-load-3-hours/ A continuous three-hour serving pass on two identical Ryzen AI Max+ 395 units at concurrency four: how far the decode pace drifted between the first five minutes and the last, and what share of requests came back complete. Sun, 16 Aug 2026 00:00:00 GMT lab_single_run https://agmind.ru/reports/answerless-http-200/ https://agmind.ru/reports/answerless-http-200/ A reasoning model can consume its entire token budget thinking and return a successful response with zero user-visible text. We measured how often that happens on two model families — and why TTFT dashboards look perfect while it does. Fri, 14 Aug 2026 00:00:00 GMT lab_repeated https://agmind.ru/reports/cyrillic-token-cost/ https://agmind.ru/reports/cyrillic-token-cost/ Our long-context corpus was cut by a chars-per-token estimate that held for English and missed Russian by a third. The per-item measured counts are published — here is what they say about planning context windows for Cyrillic, and how we caught our own mislabel. Fri, 14 Aug 2026 00:00:00 GMT lab_repeated https://agmind.ru/essays/luddites-were-right/ https://agmind.ru/essays/luddites-were-right/ A year of journal entries: every time AI lied to me and every time it genuinely came through. The verification protocol that grew out of it, and why the real Luddite story is about quality and the price of the transition, not fear of machines. Fri, 14 Aug 2026 00:00:00 GMT https://agmind.ru/reports/dspark-speed-numbers-disagree/ https://agmind.ru/reports/dspark-speed-numbers-disagree/ The same model on the same two GB10 nodes is reported anywhere from the high fifties to the low seventies of tokens per second. Every one of those figures can be correct — here is the configuration behind each, and which number to quote for which question. Thu, 13 Aug 2026 00:00:00 GMT lab_single_run https://agmind.ru/reports/concurrency-capacity-strix-halo/ https://agmind.ru/reports/concurrency-capacity-strix-halo/ A measured concurrency ladder on one Strix Halo unit: what the wait looks like at one, four and eight simultaneous chats, what stayed correct, and why 'tokens per second' cannot answer the sizing question at all. Wed, 12 Aug 2026 00:00:00 GMT lab_repeated https://agmind.ru/reports/dspark-speculative-1m-dgx-spark/ https://agmind.ru/reports/dspark-speculative-1m-dgx-spark/ Our fp8 DSpark bring-up on GB10, an independent reproduction of the community NVFP4 build, and the depth data nobody had published: the single-stream curve to a real million-token prompt — plus two negative results, stated as such. Wed, 05 Aug 2026 00:00:00 GMT lab_single_run https://agmind.ru/reports/dsv4-flash-moe-backend-dgx-spark/ https://agmind.ru/reports/dsv4-flash-moe-backend-dgx-spark/ Serving DeepSeek-V4-Flash-0731 across two DGX Spark nodes: the image ships a MoE kernel built for this model's MXFP4 experts on this silicon — and auto-selection skips it. Picking it by hand is worth double-digit percent of decode. Recipe and raw outputs are public. Mon, 03 Aug 2026 00:00:00 GMT lab_single_run https://agmind.ru/reports/model-quant-backend-strix-halo/ https://agmind.ru/reports/model-quant-backend-strix-halo/ Same Strix Halo unit, three swaps measured one at a time: a second model family reproduces the empty-answer failure mode; Q8_0 buys nothing measurable on these workloads; ROCm works on gfx1151 but decodes slower than Vulkan. Mon, 03 Aug 2026 00:00:00 GMT lab_repeated https://agmind.ru/tested-configurations/strix-halo-qwen36-llamacpp-vulkan-docsession/ https://agmind.ru/tested-configurations/strix-halo-qwen36-llamacpp-vulkan-docsession/ The second question over the same pasted document: with the server prompt cache off it pays the full half-minute prefill again; with the cache on it answers in under a second. One setting turns a document workflow from unusable into fluid. Mon, 03 Aug 2026 00:00:00 GMT lab_repeated https://agmind.ru/tested-configurations/strix-halo-qwen36-llamacpp-vulkan-endurance/ https://agmind.ru/tested-configurations/strix-halo-qwen36-llamacpp-vulkan-endurance/ A continuous 3-hour closed-loop pass at concurrency 4 on both units: all requests answered, decode pace held between the first and the last five minutes, no thermal collapse — with the two units running at visibly different die temperatures. Mon, 03 Aug 2026 00:00:00 GMT lab_single_run https://agmind.ru/tested-configurations/strix-halo-qwen36-llamacpp-vulkan-interactive/ https://agmind.ru/tested-configurations/strix-halo-qwen36-llamacpp-vulkan-interactive/ Single-user everyday chat on a Beelink GTR9 Pro: instant answers with reasoning disabled, a broken configuration at the default reasoning budget, and a verbosity limit that never fully clears. Mon, 03 Aug 2026 00:00:00 GMT lab_repeated https://agmind.ru/tested-configurations/strix-halo-qwen36-llamacpp-vulkan-longcontext/ https://agmind.ru/tested-configurations/strix-halo-qwen36-llamacpp-vulkan-longcontext/ Needle retrieval across a context ladder up to 30k measured tokens with unanswerable controls: retrieval and honesty both held everywhere measured — while time to the first token grew linearly into tens of seconds. Mon, 03 Aug 2026 00:00:00 GMT lab_repeated https://agmind.ru/tested-configurations/strix-halo-qwen36-llamacpp-vulkan-structured/ https://agmind.ru/tested-configurations/strix-halo-qwen36-llamacpp-vulkan-structured/ Deterministic extraction, classification and structured generation in strict JSON: a perfect task-success rate in both reasoning modes on this corpus — with a seventeenfold end-to-end time difference between them. Mon, 03 Aug 2026 00:00:00 GMT lab_repeated https://agmind.ru/reports/ttft-thinking-model-strix-halo/ https://agmind.ru/reports/ttft-thinking-model-strix-halo/ On a Strix Halo box the conventional TTFT figure looked excellent while most requests returned an empty answer with HTTP 200. Client-side measurement, three operating settings, evidence bundles included. Sun, 02 Aug 2026 00:00:00 GMT lab_repeated https://agmind.ru/reports/dspark-dgx-spark-1m-legacy/ https://agmind.ru/reports/dspark-dgx-spark-1m-legacy/ Bring-up of DSpark speculative decoding on GB10 (sm_121) over 200G RoCE, the 1M-token depth curve nobody had published, an independent reproduction of a community build, and two honest negative results. Fri, 31 Jul 2026 00:00:00 GMT legacy https://agmind.ru/reports/strix-halo-multislot-legacy/ https://agmind.ru/reports/strix-halo-multislot-legacy/ What one 128 GB Strix Halo mini-PC did under 32 concurrent chat streams: recipes, an 8→10 concurrency cliff in every tested configuration, and where speculative decoding stops helping. Pre-methodology evidence. Fri, 31 Jul 2026 00:00:00 GMT legacy