experimentos. — Vlad’s open notebookLocal AI experiments, measured findings, and useful failures.https://experimentos-groxaxo.pages.dev/enA smaller speech model earns its place on the iPhonehttps://experimentos-groxaxo.pages.dev/journal/speech-to-text-parakeet-iphone-2026-10-03/https://experimentos-groxaxo.pages.dev/journal/speech-to-text-parakeet-iphone-2026-10-03/Aidana transcribed the 47 matched clips 3.35 times faster, with observed English WER of 2.15% and Spanish WER of 5.37%.Sat, 03 Oct 2026 00:00:00 GMTA coding agent that knows when to ask for helphttps://experimentos-groxaxo.pages.dev/journal/agentic-workflows-omp-glm-recovery-qwen-udq4-2026-10-01/https://experimentos-groxaxo.pages.dev/journal/agentic-workflows-omp-glm-recovery-qwen-udq4-2026-10-01/Three live recovery scenarios passed, each finishing with 10/10 checks.Thu, 01 Oct 2026 00:00:00 GMTCorrective data meets a fresh gameplay gatehttps://experimentos-groxaxo.pages.dev/journal/rgba-game-learning-docs-corrective-learning/https://experimentos-groxaxo.pages.dev/journal/rgba-game-learning-docs-corrective-learning/On Ubuntu CPU, the corrected policy passed 10/10 full-kitchen trials against a same-host retraining control at 7/10.Wed, 30 Sep 2026 00:00:00 GMTWhen the retention gate says stophttps://experimentos-groxaxo.pages.dev/journal/llm-polisher-retention-replay-kd-v1-20260930/https://experimentos-groxaxo.pages.dev/journal/llm-polisher-retention-replay-kd-v1-20260930/All three registered learning rates crossed the retention floor after eight steps and rolled back.Wed, 30 Sep 2026 00:00:00 GMTA second dish, a new frozen testhttps://experimentos-groxaxo.pages.dev/journal/rgba-game-learning-docs-two-dish-curriculum/https://experimentos-groxaxo.pages.dev/journal/rgba-game-learning-docs-two-dish-curriculum/The corrected candidate passed 10/10 paired two-dish trials against a frozen baseline at 0/10.Tue, 29 Sep 2026 00:00:00 GMTEight clears are still not a promotionhttps://experimentos-groxaxo.pages.dev/journal/rgba-game-learning-docs-kitchen1-end-to-end/https://experimentos-groxaxo.pages.dev/journal/rgba-game-learning-docs-kitchen1-end-to-end/The v3 protocol measured 8/10 full-stage clears versus 0/10 for its paired baseline; the strict gate rejected it.Tue, 29 Sep 2026 00:00:00 GMTLearning to play from pixelshttps://experimentos-groxaxo.pages.dev/journal/rgba-game-learning/https://experimentos-groxaxo.pages.dev/journal/rgba-game-learning/The first-dish baseline produced six successful recorded evaluations across five starting positions, with exact replay.Tue, 29 Sep 2026 00:00:00 GMTThe reusable RGBA learning pipelinehttps://experimentos-groxaxo.pages.dev/journal/rgba-game-learning-pixel/https://experimentos-groxaxo.pages.dev/journal/rgba-game-learning-pixel/Owned frames, bounded controls, saved state, teacher-data export and exact replay form one auditable pipeline.Tue, 29 Sep 2026 00:00:00 GMTWhat counts as a successful episode?https://experimentos-groxaxo.pages.dev/journal/rgba-game-learning-docs-episode-logic-fixes/https://experimentos-groxaxo.pages.dev/journal/rgba-game-learning-docs-episode-logic-fixes/The evaluator tightened survival, terminal-state and exact-replay checks without rewriting historical experimental outcomes.Tue, 29 Sep 2026 00:00:00 GMTFinding a local model for the inboxhttps://experimentos-groxaxo.pages.dev/journal/llm-bench-2026-09-27/https://experimentos-groxaxo.pages.dev/journal/llm-bench-2026-09-27/Bonsai-2-27B PQ2_0 with MTP passed 48/50 tasks with thinking on, at a 15.8-second median on an RTX 3060.Sun, 27 Sep 2026 00:00:00 GMTTeaching a 4B model to read tmux statushttps://experimentos-groxaxo.pages.dev/journal/lora-qwen35-4b-tmux-status-qlora-2026-09-24/https://experimentos-groxaxo.pages.dev/journal/lora-qwen35-4b-tmux-status-qlora-2026-09-24/The classifier matched its teacher on 192/202 examples, with macro-F1 0.863.Thu, 24 Sep 2026 00:00:00 GMTAn image editor that respects the pixels you keephttps://experimentos-groxaxo.pages.dev/journal/comfyui-qwen-studio-2026-09-22/https://experimentos-groxaxo.pages.dev/journal/comfyui-qwen-studio-2026-09-22/A real masked object removal finished in 24.39 seconds and preserved every unselected decoded pixel exactly.Tue, 22 Sep 2026 00:00:00 GMTCharacter LoRAs: identity, motion, and continuityhttps://experimentos-groxaxo.pages.dev/journal/comfyui-lucianax-loras-2026-09-22/https://experimentos-groxaxo.pages.dev/journal/comfyui-lucianax-loras-2026-09-22/Three initial adapters completed: 600 LTX steps and two Qwen identity runs of 800 steps each.Tue, 22 Sep 2026 00:00:00 GMTFifteen pictures, one photographic-style adapterhttps://experimentos-groxaxo.pages.dev/journal/comfyui-photo15x-qwen21-2026-09-22/https://experimentos-groxaxo.pages.dev/journal/comfyui-photo15x-qwen21-2026-09-22/213 files became 97 unique decoded images; the curated random selection used 15 distinct photographs or photo-like images.Tue, 22 Sep 2026 00:00:00 GMTMaking the keyboard lock key really sleep the phonehttps://experimentos-groxaxo.pages.dev/journal/motorola-g54-findings/https://experimentos-groxaxo.pages.dev/journal/motorola-g54-findings/A keyboard-specific handler preserved the lock gesture and the display reached OFF roughly 0.7 seconds later.Tue, 22 Sep 2026 00:00:00 GMTMasked cleanup before a fresh LoRAhttps://experimentos-groxaxo.pages.dev/journal/comfyui-susanah-qwen21-2026-09-22/https://experimentos-groxaxo.pages.dev/journal/comfyui-susanah-qwen21-2026-09-22/25 distinct images were cleaned into lossless copies, preserving dimensions and all decoded pixels outside each mask.Tue, 22 Sep 2026 00:00:00 GMTMoving speech decode onto the GPUhttps://experimentos-groxaxo.pages.dev/journal/moss-tts-awq-fp8-codec-deploy-2026-09-22/https://experimentos-groxaxo.pages.dev/journal/moss-tts-awq-fp8-codec-deploy-2026-09-22/Decode dropped from 3.36 seconds to 0.24; total render wall time dropped from 35.1 to 24.4 seconds.Tue, 22 Sep 2026 00:00:00 GMTPutting a watermark gate in front of traininghttps://experimentos-groxaxo.pages.dev/journal/comfyui-susanah-qwen21-screened-v2-2026-09-22/https://experimentos-groxaxo.pages.dev/journal/comfyui-susanah-qwen21-screened-v2-2026-09-22/The binary screen accepted 24/25 cleaned images; both known positive controls were detected.Tue, 22 Sep 2026 00:00:00 GMTA complete image stack on one 24GB GPUhttps://experimentos-groxaxo.pages.dev/journal/qwen21-te-awq/https://experimentos-groxaxo.pages.dev/journal/qwen21-te-awq/Eight matched 1024-square scenes averaged 93.352 seconds with FP8 and 46.219 with opt-in INT8.Mon, 21 Sep 2026 00:00:00 GMTLong speech, bounded memoryhttps://experimentos-groxaxo.pages.dev/journal/moss-tts-hq4-moss54d-autochunk-2026-09-21/https://experimentos-groxaxo.pages.dev/journal/moss-tts-hq4-moss54d-autochunk-2026-09-21/A 551-character, eight-sentence input rendered in two chunks with 48.0 seconds of healthy audio.Mon, 21 Sep 2026 00:00:00 GMTRound-trip tests expose a broken encoderhttps://experimentos-groxaxo.pages.dev/journal/moss-tts-awq-codec-roundtrip-fp8-2026-09-21/https://experimentos-groxaxo.pages.dev/journal/moss-tts-awq-codec-roundtrip-fp8-2026-09-21/The new 1.76GB FP8-E4M3 codec reached FP32-control parity in the measured round-trip test with its encoder intact.Mon, 21 Sep 2026 00:00:00 GMTThe clone bug was in the reference encoderhttps://experimentos-groxaxo.pages.dev/journal/moss-tts-awq-clone-fp32-reference-2026-09-21/https://experimentos-groxaxo.pages.dev/journal/moss-tts-awq-clone-fp32-reference-2026-09-21/Using FP32 reference encoding lifted AWQ WavLM target similarity from 0.8475 to 0.9556 while retaining INT8 output decode.Mon, 21 Sep 2026 00:00:00 GMTThe faster image pipeline that produced nothinghttps://experimentos-groxaxo.pages.dev/journal/qwen21-2gpu-broken/https://experimentos-groxaxo.pages.dev/journal/qwen21-2gpu-broken/The balanced two-GPU path took 48.1 seconds but every image was invalid flat gray noise.Mon, 21 Sep 2026 00:00:00 GMTThe same story, six precision combinationshttps://experimentos-groxaxo.pages.dev/journal/moss-tts-awq-quant-matrix-stories-2026-09-21/https://experimentos-groxaxo.pages.dev/journal/moss-tts-awq-quant-matrix-stories-2026-09-21/Generator precision affected the measured ranking; changing the output codec shifted similarity by at most about 0.001.Mon, 21 Sep 2026 00:00:00 GMTBalancing GPU experts when the model will not fithttps://experimentos-groxaxo.pages.dev/journal/qwen-next-flash-udq4kxl-cpu-moe-200k-2026-09-20/https://experimentos-groxaxo.pages.dev/journal/qwen-next-flash-udq4kxl-cpu-moe-200k-2026-09-20/An explicit 12/12/12 GPU expert split plus 12 CPU blocks measured 31.1 tokens per second.Sun, 20 Sep 2026 00:00:00 GMTPipeline parallelism, INT4, and a 262K contexthttps://experimentos-groxaxo.pages.dev/journal/qwen-next-flash-vllm-int4-autoround-3x3090-2026-09-20/https://experimentos-groxaxo.pages.dev/journal/qwen-next-flash-vllm-int4-autoround-3x3090-2026-09-20/The patched vLLM stack measured 67.4 tokens per second for one stream and 73.1 aggregate for two.Sun, 20 Sep 2026 00:00:00 GMTA speech service that sleeps between requestshttps://experimentos-groxaxo.pages.dev/journal/moss-tts-hq4-int8-codec-lazy-server-2026-09-19/https://experimentos-groxaxo.pages.dev/journal/moss-tts-hq4-int8-codec-lazy-server-2026-09-19/The recorded service used about 10.3GB resident memory; first render took 14.8 seconds and warm render 5.7.Sat, 19 Sep 2026 00:00:00 GMTA quantization job goes silenthttps://experimentos-groxaxo.pages.dev/journal/moss-tts-hq4-phase1-incident-2026-09/https://experimentos-groxaxo.pages.dev/journal/moss-tts-hq4-phase1-incident-2026-09/The full-spec job stopped after six of 36 blocks; the monitor detected the dead wrapper and idle GPU.Fri, 18 Sep 2026 00:00:00 GMTAWQ speech: fast enough, with a memory catchhttps://experimentos-groxaxo.pages.dev/journal/moss-tts-awq-awq-3090-2026-09/https://experimentos-groxaxo.pages.dev/journal/moss-tts-awq-awq-3090-2026-09/AWQ passed 33/33 prompts with pooled real-time factor 0.43 and 35.8 tokens per second.Fri, 18 Sep 2026 00:00:00 GMTDo you really need the bigger embedding model?https://experimentos-groxaxo.pages.dev/journal/nemotron-embed-1b-vs-8b-awq-2026-09-18/https://experimentos-groxaxo.pages.dev/journal/nemotron-embed-1b-vs-8b-awq-2026-09-18/The 1B encoder embedded 1,004 documents per second; the 8B encoder reached 181, with no proven retrieval-quality win.Fri, 18 Sep 2026 00:00:00 GMTFP32, BF16, INT8: separating codec drift from speechhttps://experimentos-groxaxo.pages.dev/journal/moss-tts-hq4-codec-abc-2026-09/https://experimentos-groxaxo.pages.dev/journal/moss-tts-hq4-codec-abc-2026-09/Pure TTS legs shared token sequences, and normalized transcription differences were zero across the tested codec legs.Fri, 18 Sep 2026 00:00:00 GMTKeeping the audio-sensitive layers intacthttps://experimentos-groxaxo.pages.dev/journal/moss-tts-hq4-phase0-fidelity-envelope-2026-09/https://experimentos-groxaxo.pages.dev/journal/moss-tts-hq4-phase0-fidelity-envelope-2026-09/The HQ4 preview matched 96.5% of text tokens in the greedy test and passed three staged-memory probes at 6.7–6.8 GiB.Fri, 18 Sep 2026 00:00:00 GMTStaging the speech stack under a 12GB envelopehttps://experimentos-groxaxo.pages.dev/journal/moss-tts-hq4-pr1-hardening-validation-2026-09-18/https://experimentos-groxaxo.pages.dev/journal/moss-tts-hq4-pr1-hardening-validation-2026-09-18/The staged run peaked at 6.63 GiB, down from 13.01 before staging; 252 INT4 projections passed the verifier.Fri, 18 Sep 2026 00:00:00 GMTThree draft tokens hit the sweet spothttps://experimentos-groxaxo.pages.dev/journal/qwen-next-flash-exl3-350hq-mtp-draft-q8-2026-09-16/https://experimentos-groxaxo.pages.dev/journal/qwen-next-flash-exl3-350hq-mtp-draft-q8-2026-09-16/Draft length three reached a short-prose mean of 88.16 tokens per second, versus 64.67 with drafting off.Wed, 16 Sep 2026 00:00:00 GMTFrom a model download to a working local servicehttps://experimentos-groxaxo.pages.dev/journal/qwen-next-flash/https://experimentos-groxaxo.pages.dev/journal/qwen-next-flash/The deployment notebook records the EXL3 serving family, cache choices and real runtime constraints.Tue, 15 Sep 2026 00:00:00 GMTThe EXL3 quantization journeyhttps://experimentos-groxaxo.pages.dev/journal/qwen-next-flash-exl3-350hq-375-2026-09-15/https://experimentos-groxaxo.pages.dev/journal/qwen-next-flash-exl3-350hq-375-2026-09-15/The notebook follows official BF16 weights through EXL3 conversion, artifact checks and deployment work.Tue, 15 Sep 2026 00:00:00 GMTA model proxy earns its promotionhttps://experimentos-groxaxo.pages.dev/journal/qwen38-proxy-pr5-hardening-validation-2026-09-14/https://experimentos-groxaxo.pages.dev/journal/qwen38-proxy-pr5-hardening-validation-2026-09-14/The promoted merge passed 107 tests with one environmental skip.Mon, 14 Sep 2026 00:00:00 GMTThe best checkpoint was not the last onehttps://experimentos-groxaxo.pages.dev/journal/breeze-tts2-latam-es-mf-polish-2026-09/https://experimentos-groxaxo.pages.dev/journal/breeze-tts2-latam-es-mf-polish-2026-09/Polish step 400 measured 6.1% word error rate; the final step 1500 measured 42.9%.Fri, 11 Sep 2026 00:00:00 GMTRebuilding calibration for an agentic quanthttps://experimentos-groxaxo.pages.dev/journal/qwen38-flash-next-unc-q4km-agentic-v2-2026-09-08/https://experimentos-groxaxo.pages.dev/journal/qwen38-flash-next-unc-q4km-agentic-v2-2026-09-08/The clean agentic importance matrix recovered the hard T3 task; later draft decoding improved measured throughput.Tue, 08 Sep 2026 00:00:00 GMTA draft head, a rollback fix, and a midnight incidenthttps://experimentos-groxaxo.pages.dev/journal/qwen38-flash-next-unc-mtp-draft-2026-09-07/https://experimentos-groxaxo.pages.dev/journal/qwen38-flash-next-unc-mtp-draft-2026-09-07/The rollback fix brought draft decoding to parity at temperature 0.7 and a 24% gain at greedy sampling.Mon, 07 Sep 2026 00:00:00 GMTMaking room for a 200K contexthttps://experimentos-groxaxo.pages.dev/journal/qwen38-flash-next-udq3xl-3x3090-200k-2026-08-29/https://experimentos-groxaxo.pages.dev/journal/qwen38-flash-next-udq3xl-3x3090-200k-2026-08-29/The warm proxy decode median was 59.17 tokens per second; a 198,960-token request passed with GPU headroom.Sat, 29 Aug 2026 00:00:00 GMTDFlash2 on a single RTX 3090https://experimentos-groxaxo.pages.dev/journal/llamacpp-docs-experiments-ornith-dflash2-rtx3090/https://experimentos-groxaxo.pages.dev/journal/llamacpp-docs-experiments-ornith-dflash2-rtx3090/Ornith measured 210.16 tokens per second on the realistic 256-token prompt and 270.40 on predictable prose.Thu, 27 Aug 2026 00:00:00 GMTLess thinking, fewer tokens: a template A/Bhttps://experimentos-groxaxo.pages.dev/journal/qwen38-templates-2026-08-27-medium-xhigh-template-ab/https://experimentos-groxaxo.pages.dev/journal/qwen38-templates-2026-08-27-medium-xhigh-template-ab/Sharp reduced completion tokens by 20.6%, 20.1% and 16.0% across the three deployments.Thu, 27 Aug 2026 00:00:00 GMTThe real cost of a five-second videohttps://experimentos-groxaxo.pages.dev/journal/comfyui-minimaxh3/https://experimentos-groxaxo.pages.dev/journal/comfyui-minimaxh3/At 960×544, dual INT8 completed a 124-frame, 20-step run in 354.580 seconds.Thu, 27 Aug 2026 00:00:00 GMT