https://1bit.monster/1bit-blog.html How the 1-bit inference engine actually gets built — NPU, ROCm, kernels, formats. en Thu, 24 Sep 2026 10:30:09 +0000 https://1bit.monster/1bit-post-lemonade-v1190.html https://1bit.monster/1bit-post-lemonade-v1190.html Wed, 02 Sep 2026 00:00:00 +0000 The embedded Lemonade server core is re-vendored to v11.9.0 — allowed_origins/CORS hardening, llama.cpp to b10723, the streaming-stall fix, and directory-watcher + extra-model classification fixes. We retired stream_stall_timeout and kept the newer hrx-b66 pin; everything else now matches upstream. https://1bit.monster/1bit-post-zaya-q4nx-hrx.html https://1bit.monster/1bit-post-zaya-q4nx-hrx.html Mon, 31 Aug 2026 00:00:00 +0000 Milestone: an 8-billion-parameter model now runs entirely on AMD's laptop NPU — attention and experts on-device, served over HTTP, decode 8.4 tok/s. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/research/ws12-hrx-loom/README.md https://1bit.monster/1bit-post-lemonade-v1181.html https://1bit.monster/1bit-post-lemonade-v1181.html Thu, 27 Aug 2026 00:00:00 +0000 The embedded Lemonade server core is re-vendored to v11.8.1 — upstream sync landed; see the post for what changed. Engine HF coverage: 569 architecture tokens, 2,030 HF arch strings, 326,992/327,349 checkpoints mapped. https://1bit.monster/1bit-post-hrx-engine.html https://1bit.monster/1bit-post-hrx-engine.html Sat, 29 Aug 2026 00:00:00 +0000 Milestone: AMD's experimental NPU runtime now runs inside our engine — one process, two runtimes, roughly twice the warm decode, with the honest ceiling measured. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/docs/research/hrx-backend.md https://1bit.monster/1bit-post-lemonade-v1180.html https://1bit.monster/1bit-post-lemonade-v1180.html Thu, 27 Aug 2026 00:00:00 +0000 Milestone: the AMD SDK release built for our chip — the 15th backend (DeepSeek V4 Flash for Strix Halo) embedded, plus download resume and rate limits. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/docs/guides/Lemonade-Compat.md https://1bit.monster/1bit-post-1775-models.html https://1bit.monster/1bit-post-1775-models.html Tue, 25 Aug 2026 00:00:00 +0000 One C++ binary, 569 architecture tokens, 2,030 HF arch strings, 99.89% HuggingFace coverage, 326,992 checkpoints mapped. Every model you can download. https://1bit.monster/1bit-post-local-rag.html https://1bit.monster/1bit-post-local-rag.html Tue, 25 Aug 2026 00:00:00 +0000 Milestone: fully local RAG on one binary — embed, index, retrieve and generate on your own hardware, no cloud, no Python. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/docs/CODEBASE.md https://1bit.monster/1bit-post-1bit-explained.html https://1bit.monster/1bit-post-1bit-explained.html Tue, 25 Aug 2026 00:00:00 +0000 Milestone: models packed to ~1.6 bits per weight run on hardware you already own — memory drops ~16x and matmuls become additions. What 1-bit actually does to a model, explained plainly. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/docs/journey.md https://1bit.monster/1bit-post-rocm-vs-cuda.html https://1bit.monster/1bit-post-rocm-vs-cuda.html Tue, 25 Aug 2026 00:00:00 +0000 Milestone: 1-bit matmuls on AMD silicon measured against CUDA from real kernel work — the comparison behind why every model runs on AMD's stack. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/docs/engine_comparison_report.md https://1bit.monster/1bit-post-lemonade-v1170.html https://1bit.monster/1bit-post-lemonade-v1170.html Mon, 24 Aug 2026 00:00:00 +0000 Milestone: the SDK sync loop — Lemonade v11.7.0's core embedded so the engine stays current with every AMD release, 14 backends included. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/docs/guides/Lemonade-Compat.md https://1bit.monster/1bit-post-fleet.html https://1bit.monster/1bit-post-fleet.html Sat, 22 Aug 2026 00:00:00 +0000 Milestone: a two-computer AI fleet with zero cloud — 35B and 8B models served locally across a LAN, both instances fed entirely by our own engine. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/docs/journey.md https://1bit.monster/1bit-post-qwen36.html https://1bit.monster/1bit-post-qwen36.html Sat, 22 Aug 2026 00:00:00 +0000 Milestone: a 35-billion-parameter model streaming on a laptop NPU — AMD's own numbers decoded and independently verified, plus our first Vulkan pass. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/docs/journey.md https://1bit.monster/1bit-post-flm-zoo.html https://1bit.monster/1bit-post-flm-zoo.html Sat, 22 Aug 2026 00:00:00 +0000 Milestone: AMD's entire NPU model collection extracted — 37 models and 209 bitstreams, auto-detected from the model files with zero configuration. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/docs/research/fastflowlm-analysis/README.md https://1bit.monster/1bit-post-zyphra.html https://1bit.monster/1bit-post-zyphra.html Sat, 22 Aug 2026 00:00:00 +0000 Milestone: Zyphra's entire model family runs at 1-bit precision on GPU — all four models validated end to end, including a reasoning model at 26 tok/s. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/docs/model-families/zyphra.md https://1bit.monster/1bit-post-binary-formats.html https://1bit.monster/1bit-post-binary-formats.html Sat, 22 Aug 2026 00:00:00 +0000 Milestone: one binary runs every model format — 1-bit, ternary and binary checkpoints on NPU and GPU alike, no rebuilds, no per-model config. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/docs/journey.md https://1bit.monster/1bit-post-unsloth.html https://1bit.monster/1bit-post-unsloth.html Sat, 22 Aug 2026 00:00:00 +0000 Milestone: training and fine-tuning land on AMD hardware — 2x faster, 70% less memory, fits in 3 GB, alongside our inference engine. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/docs/wiki/unsloth-dynamic-ggufs.md https://1bit.monster/1bit-post-tokenrouter.html https://1bit.monster/1bit-post-tokenrouter.html Sat, 22 Aug 2026 00:00:00 +0000 Milestone: token-level routing between NPU and GPU — one chip drafts, another verifies, and real logprobs ride the wire between them. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/docs/journey.md https://1bit.monster/1bit-post-halo-reviews.html https://1bit.monster/1bit-post-halo-reviews.html Sat, 22 Aug 2026 00:00:00 +0000 Milestone: the 50-TOPS laptop NPU driven directly — what the reviewers missed, in our own measurements. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/docs/journey.md https://1bit.monster/1bit-post-dspark.html https://1bit.monster/1bit-post-dspark.html Sat, 22 Aug 2026 00:00:00 +0000 Milestone: a speculative-decoding method that promised 572 tok/s delivered ~0.1-0.2 tok/s — disproven with measurements, and what it would take to fix. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/spec-decode/README.md https://1bit.monster/1bit-post-three-bugs.html https://1bit.monster/1bit-post-three-bugs.html Sat, 22 Aug 2026 00:00:00 +0000 Milestone: full-speed NPU decode that was secretly garbage, made correct — three silent bugs found and fixed at 97 tok/s. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/docs/journey.md https://1bit.monster/1bit-post-one-engine.html https://1bit.monster/1bit-post-one-engine.html Sat, 22 Aug 2026 00:00:00 +0000 Milestone: one binary runs every model on every chip — architecture auto-detected from the file, backend auto-chosen, no per-model builds. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/docs/CODEBASE.md https://1bit.monster/1bit-post-sprint.html https://1bit.monster/1bit-post-sprint.html Sat, 22 Aug 2026 00:00:00 +0000 Milestone: NPU inference from 1,930 to 3.4 ms per token in four days — a 72x speedup through batch decode, fused dispatch and INT8 kernels. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/engine/npu/BENCHMARKS.md https://1bit.monster/1bit-post-50t.html https://1bit.monster/1bit-post-50t.html Sat, 22 Aug 2026 00:00:00 +0000 Milestone: the first month's NPU benchmarks, kept honestly — 50 TOPS on a laptop, early numbers flagged as unsourced history. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/engine/npu/BENCHMARKS.md https://1bit.monster/1bit-post-fused.html https://1bit.monster/1bit-post-fused.html Sat, 22 Aug 2026 00:00:00 +0000 Milestone: a model's layers fused into single NPU calls — six times fewer dispatches per layer, making the laptop NPU a practical inference target. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/npu-infer/docs/fusion-level-0.md https://1bit.monster/1bit-post-npu-reversal.html https://1bit.monster/1bit-post-npu-reversal.html Sat, 22 Aug 2026 00:00:00 +0000 Milestone: AMD's locked NPU stack replaced with open C++ in four days — a laptop, a disassembler, 22 proprietary libraries, zero documentation. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/docs/research/fastflowlm-analysis/README.md https://1bit.monster/1bit-post-coding-agent.html https://1bit.monster/1bit-post-coding-agent.html Sat, 22 Aug 2026 00:00:00 +0000 Milestone: a coding agent that runs entirely on your hardware — a CLI with NPU-native inference, package management, extensions and skills, in the monorepo. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/docs/archive/openspec/changes/1bit-pi-clone/proposal.md https://1bit.monster/1bit-post-npu-v12.html https://1bit.monster/1bit-post-npu-v12.html Sat, 22 Aug 2026 00:00:00 +0000 Milestone: NPU batch decode hit 97 tok/s — a 24x speedup in one session (244 to 10 ms/token), 46% faster than AMD's own stack. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/engine/npu/BENCHMARKS.md https://1bit.monster/1bit-post-historical.html https://1bit.monster/1bit-post-historical.html Sat, 22 Aug 2026 00:00:00 +0000 Milestone: an early two-lane local run, kept for the record — one model, iGPU and NPU, both green. Full technical write-up: https://github.com/1bit-MONSTER/1bit-MONSTER/blob/main/engine/npu/BENCHMARKS.md