<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>Haku Lab</title><description>Short, honest write-ups of experiments with small language models that run on-device.</description><link>https://lab.haku.sh/</link><language>en</language><item><title>Slice adapters vs LoRA on a ternary model</title><link>https://lab.haku.sh/slice-adapters-vs-lora/</link><guid isPermaLink="true">https://lab.haku.sh/slice-adapters-vs-lora/</guid><description>Retrain one slice of a ternary model&apos;s feed-forward channels in place, instead of adding a LoRA side path. Compared on four ternary models, timed on the M1 Neural Engine.</description><pubDate>Mon, 09 Nov 2026 00:00:00 GMT</pubDate><category>ternary</category><category>LoRA</category><category>fine-tuning</category><category>adapters</category><category>ANE</category><category>MoE</category><category>BitNet</category><category>on-device</category></item><item><title>Fixing small models by editing their weights: 5 of 7 failures</title><link>https://lab.haku.sh/weight-edit/</link><guid isPermaLink="true">https://lab.haku.sh/weight-edit/</guid><description>An MIT team fixed a looping model by editing its weights, with no training. We tried the same idea on 7 failures in 6 of our small models, with a random edit of the same size as the control.</description><pubDate>Tue, 03 Nov 2026 00:00:00 GMT</pubDate><category>weight-editing</category><category>interpretability</category><category>ternary</category><category>MoE</category><category>evaluation</category><category>on-device</category></item><item><title>An experiment to run LiquidAI&apos;s DSpark drafter on the Neural Engine</title><link>https://lab.haku.sh/ane-drafter/</link><guid isPermaLink="true">https://lab.haku.sh/ane-drafter/</guid><description>The DSpark draft model for LFM2.5-1.2B, moved onto the Apple Neural Engine while Metal verifies. Everything runs on-device.</description><pubDate>Fri, 21 Aug 2026 00:00:00 GMT</pubDate><category>ANE</category><category>speculative-decoding</category><category>LFM2</category><category>Core ML</category><category>on-device</category><category>inference</category></item><item><title>Does the 27B Bonsai actually deliver 27B?</title><link>https://lab.haku.sh/ternary-27b-agentic-benchmark/</link><guid isPermaLink="true">https://lab.haku.sh/ternary-27b-agentic-benchmark/</guid><description>Prism shipped a 27B that fits a laptop. We put it through a battery of agentic traps to see if it reasons like its size.</description><pubDate>Wed, 15 Jul 2026 00:00:00 GMT</pubDate><category>ternary</category><category>evaluation</category><category>tool-calls</category><category>BitNet</category><category>on-device</category></item><item><title>Running a ternary 27B on a single A10</title><link>https://lab.haku.sh/ternary-27b-single-a10-gpu/</link><guid isPermaLink="true">https://lab.haku.sh/ternary-27b-single-a10-gpu/</guid><description>A 27B model whose weights are all −1, 0 or +1 — what it takes to run, and how fast it goes.</description><pubDate>Wed, 15 Jul 2026 00:00:00 GMT</pubDate><category>ternary</category><category>on-device</category><category>inference</category><category>BitNet</category><category>GPU</category></item><item><title>Fine-tuning a 230M model to know when you&apos;ve stopped talking</title><link>https://lab.haku.sh/end-of-turn-detection-lfm2/</link><guid isPermaLink="true">https://lab.haku.sh/end-of-turn-detection-lfm2/</guid><description>A from-scratch 5M-parameter classifier plateaued at 30% recall on real speech. A LoRA fine-tune of a 230M pretrained model, on the same data, didn&apos;t.</description><pubDate>Thu, 02 Jul 2026 00:00:00 GMT</pubDate><category>fine-tuning</category><category>LoRA</category><category>on-device</category><category>specialist</category><category>voice</category></item><item><title>Indexing code by parsing it, not chunking it</title><link>https://lab.haku.sh/codesearch/</link><guid isPermaLink="true">https://lab.haku.sh/codesearch/</guid><description>Haku&apos;s code search reads each language&apos;s syntax tree, indexes real symbols, and embeds them on the Neural Engine — running inside the terminal, nothing leaving the Mac.</description><pubDate>Thu, 18 Jun 2026 00:00:00 GMT</pubDate><category>code-search</category><category>embeddings</category><category>ANE</category><category>on-device</category><category>tree-sitter</category></item><item><title>Resolving natural-language dates with a tiny specialist model</title><link>https://lab.haku.sh/datelm/</link><guid isPermaLink="true">https://lab.haku.sh/datelm/</guid><description>Division of labour — a 14M model maps language to structure; ordinary code does the arithmetic</description><pubDate>Thu, 18 Jun 2026 00:00:00 GMT</pubDate><category>dates</category><category>structured</category><category>on-device</category><category>specialist</category><category>parsing</category></item><item><title>A 9M ternary encoder for accessibility-tree elements</title><link>https://lab.haku.sh/domlm/</link><guid isPermaLink="true">https://lab.haku.sh/domlm/</guid><description>One layer of the filter between raw accessibility data and the agent — element routing, on the Neural Engine.</description><pubDate>Thu, 18 Jun 2026 00:00:00 GMT</pubDate><category>ternary</category><category>encoder</category><category>ANE</category><category>accessibility</category><category>routing</category><category>on-device</category></item><item><title>An experiment to format dictation in under 100 milliseconds</title><link>https://lab.haku.sh/formatter/</link><guid isPermaLink="true">https://lab.haku.sh/formatter/</guid><description>A 60M-parameter T5, distilled from a larger model, that cleans up raw speech on-device — split across the Neural Engine and the GPU.</description><pubDate>Thu, 18 Jun 2026 00:00:00 GMT</pubDate><category>T5</category><category>distillation</category><category>ANE</category><category>GPU</category><category>on-device</category><category>dictation</category></item><item><title>Grammar correction with a diffusion LoRA on a frozen model</title><link>https://lab.haku.sh/gec/</link><guid isPermaLink="true">https://lab.haku.sh/gec/</guid><description>A small adapter turns a frozen left-to-right model into a one-pass, both-directions corrector.</description><pubDate>Thu, 18 Jun 2026 00:00:00 GMT</pubDate><category>diffusion</category><category>LoRA</category><category>grammar-correction</category><category>frozen-model</category><category>bidirectional</category><category>on-device</category></item><item><title>An experiment to cut down TTS latency</title><link>https://lab.haku.sh/onepass-tts/</link><guid isPermaLink="true">https://lab.haku.sh/onepass-tts/</guid><description>Making a text-to-speech model decode in one pass instead of sixteen, so it runs fast on-device. With audio you can play.</description><pubDate>Thu, 18 Jun 2026 00:00:00 GMT</pubDate><category>TTS</category><category>on-device</category><category>LoRA</category><category>parallel-decode</category><category>audio</category><category>Qwen3-TTS</category></item><item><title>Parallel decoding of audio tokens in a text-to-speech model</title><link>https://lab.haku.sh/parallel-tts/</link><guid isPermaLink="true">https://lab.haku.sh/parallel-tts/</guid><description>Decoding 16 residual codebooks in one pass instead of sixteen. On-device.</description><pubDate>Thu, 18 Jun 2026 00:00:00 GMT</pubDate><category>TTS</category><category>parallel-decode</category><category>on-device</category><category>codebooks</category><category>evaluation</category></item><item><title>Diffusion-draft speculative decoding: an eight-prompt probe</title><link>https://lab.haku.sh/specdecode/</link><guid isPermaLink="true">https://lab.haku.sh/specdecode/</guid><description>A diffusion model drafts, an autoregressive model verifies. Acceptance measured on eight prompts.</description><pubDate>Thu, 18 Jun 2026 00:00:00 GMT</pubDate><category>diffusion</category><category>speculative-decoding</category><category>autoregressive</category><category>inference</category><category>evaluation</category></item><item><title>Structured tool-calls from a 27M model</title><link>https://lab.haku.sh/toolcall/</link><guid isPermaLink="true">https://lab.haku.sh/toolcall/</guid><description>How small can a model be and still learn a structured capability?</description><pubDate>Thu, 18 Jun 2026 00:00:00 GMT</pubDate><category>tool-calls</category><category>ternary</category><category>structured</category><category>on-device</category><category>fine-tuning</category><category>generalization</category></item><item><title>A few months of ternary experiments</title><link>https://lab.haku.sh/ternary-experiments/</link><guid isPermaLink="true">https://lab.haku.sh/ternary-experiments/</guid><description>Where ternary {−1, 0, +1} weights are the right tool for small, on-device models — and where they aren&apos;t.</description><pubDate>Mon, 01 Jun 2026 00:00:00 GMT</pubDate><category>ternary</category><category>QAT</category><category>on-device</category><category>BitNet</category><category>ANE</category><category>from-scratch</category></item></channel></rss>