Bonsai-27B-mlx-1bit

Full 27B-class reasoning in binary transformer weights — the first 27B-class model to run on a phone > **~14.2x** smaller than FP16 | **~90%** of FP16 intelligence retained | **~11 tok/s** on iPhone 17 Pro Max - **~3.9 GB** deployed footprint (down from ~54 GB FP16) — fits within the per-app memory budget of a high-end phone such as the iPhone 17 Pro Max - **Retains thinking, reasoning, and agentic behavior** deep in the sub-4-bit regime, where conventional low-bit representations collapse — 76.11 average across 15 thinking-mode benchmarks (89.5% of FP16), including math at 91.66 and coding at 81.88 - **End-to-end binary language weights** across embeddings, attention projections, MLP projections, and LM head, at a *true* 1.125 bits per weight — no high-precision escape hatches behind a low-bit label; the vision tower ships in compact 4-bit HQQ - **262K-token context** on-device, kept practical by the Qwen3.6-27B hybrid-attention backbone (~75% linear attention) and 4-bit KV-cache quantization - **First interactive 27B-class generation on a phone**: ~11 tok/s on iPhone 17 Pro Max; ~44 tok/s on an Apple M5 Pro laptop - **Custom 1-bit hybrid-attention kernels** on Apple MLX (Python, Swift) and CUDA — packed weights are consumed directly, never expanded back to FP16 - **Ships with a DSpark speculative-decoding drafter layer** trained against the Bonsai 27B target — a lossless **1.37x** decode speedup on the CUDA serving path - **Ternary companion**: also available as [Ternary Bonsai 27B](https://huggingface.co/prism-ml/Ternary-Bonsai-27B-mlx-2bit), the quality-oriented operating point (~7.2 GB, 95% of FP16) for laptops and GPUs - **[Whitepaper](https://github.com/PrismML-Eng/Bonsai-demo/blob/main/bonsai-27b-whitepaper.pdf)** — full methodology, benchmarks, and measurement notes - **[Demo & examples](https://github.com/PrismML-Eng/Bonsai-demo)** — serving, benchmarking, and integrating Bonsai - **Low-bit kernels**: [MLX fork](https://github.com/PrismML-Eng/mlx) (Apple Silicon) · [mlx-swift fork](https://github.com/PrismML-Eng/mlx-swift) (iOS/macOS) · [llama.cpp fork](https://github.com/PrismML-Eng/llama.cpp) (CUDA) - **[Discord](https://discord.gg/prismml)** — join the community for support, discussion, and updates Each weight is a single sign bit: `0` maps to `−scale`, `1` maps to `+scale`. Every group of 128 weights shares one FP16 scale factor. Effective bits per weight: **1.125** (1 sign bit + 16-bit scale amortized over 128 weights) — an idealized 14.2x reduction vs FP16. This is the most aggressive operating point in the Bonsai 27B family: it minimizes both stored footprint and the weight traffic incurred at every decoding step, and it is what brings a 27B model within a phone's memory budget. The binary build deploys at its native 1-bit g128 layout, so ideal and deployed sizes match. The deployed figure describes the language model alone — the only component that must stay resident for text inference; a negligible tail of normalization and scale parameters remains in higher precision. Unlike conventional low-bit builds — whose advertised labels understate their true average bit-width (a widely-used "2-bit" build of Qwen3.6-27B is really 2.8 bits/weight at 9.4 GB) — the Bonsai representation carries a bit-width that matches its name. Two optional components ship alongside the language model (on-disk sizes from the GGUF packaging): The vision tower is usually offloaded: it sits outside the accelerator's resident budget and is loaded only when an image actually arrives, so text-only serving never pays for it. The published MLX pack measures **5.13 GB** on disk (safetensors, vision tower bundled in the same file). MLX's grouped low-bit format stores both a scale and a bias per group; Bonsai's scale-only weights are packed by setting `s_mlx = 2·s_g` and `bias = −s_g`, which reproduces ±s_g exactly — the bias carries no new information — but stores two FP16 values per group where the native format stores one. The effective rate is therefore 1.25 bits/weight versus the native 1.125. This is a current MLX limitation; once MLX supports scale-only group formats, the MLX pack matches the native rate. What a device must actually accommodate is *peak* memory — weights plus KV cache plus activations and runtime buffers (~1.3 GB across backends). Measured, language model only, no KV-cache compression (sizes in decimal GB; the Q4_K_XL row is derived from its weight footprint plus the same measured cache-and-overhead build-up, all other rows directly measured): The 1-bit build holds a **100K-token context at 11.6–12.2 GB without any KV-cache compression** — a budget that fits mainstream laptops outright; the conventional Q4_K_XL build needs ~25.6 GB before the first long document is loaded. These peaks are the conservative case, with the cache left at FP16. Enabling the 4-bit KV cache shrinks the context-dependent term ~4x: the 1-bit build's 100K peak drops to ~6.8 GB, and the full 262K window fits in ~9.4 GB peak. These are the settings used for all reported benchmark results (thinking mode). You can use a simple system prompt such as: You are a helpful assistant Use the [Bonsai-demo repo](https://github.com/PrismML-Eng/Bonsai-demo) to run 1-bit Bonsai 27B on MacBook. Use the [mlx-swift fork](https://github.com/PrismML-Eng/mlx-swift) to run 1-bit Bonsai 27B on iPhone and iPad. At ~3.9 GB, the model fits within the per-app memory budget of an iPhone 17 Pro Max with headroom for the KV cache and activations. CUDA inference uses the GGUF pack of the same weights: fused 1-bit GEMM kernels in our [llama.cpp fork](https://github.com/PrismML-Eng/llama.cpp) unpack the sign codes and apply the group-wise scales inside the matrix multiply — the model is never expanded into a dense FP16 tensor in memory. See **[Bonsai-27B-gguf](https://huggingface.co/prism-ml/Bonsai-27B-gguf)** for the full CUDA, Metal, and server quickstarts. `tg128` is token-generation throughput over 128 generated tokens (the memory-bandwidth-bound, interactive phase); `pp512` is prompt-processing throughput over 512 input tokens (the compute-bound phase). Both in tokens/s. Laptop and GPU rows are measured with llama.cpp (Metal/CUDA, custom low-bit kernels) on the GGUF pack of the same weights; the iPhone row uses the MLX Swift runtime. ¹ On the phone, decode is thermally limited: sustained generation settles slightly below the cold peak — a continuous battery-drain run sustained 10.8 tok/s and generated 672 tokens per 1% of battery (roughly 67,000 tokens on a full charge). On the edge platforms the FP16 baseline (~54 GB) and even conventional "4-bit" builds (17.6 GB) do not fit at all — the meaningful statement is not a speedup ratio but that a 27B model runs on the device in the first place. The H100 row is the exception that proves the rule: at batch size 1 a datacenter GPU is limited by kernel-launch and synchronization latency rather than weight bandwidth, so the binary and ternary variants converge there (104.8 vs 98 tok/s) despite their ~1.9x difference in bytes per step. Decode energy on the M5 Pro measures **0.275 mWh/token** (with the DSpark drafter enabled) — an order of magnitude more energy-efficient per token than datacenter GPUs (0.63–1.32 mWh/token across the GPU classes). Local inference is not just private and low-latency but cheap in energy. 1-bit Bonsai 27B ships with a **DSpark** drafter layer trained against the low-bit target — a semi-autoregressive drafter with confidence-scheduled verification. Speculative decoding is lossless: verification preserves the target distribution exactly, so accepted tokens are indistinguishable from ordinary generation. The drafter is a compact **six-layer block-parallel transformer** conditioned on hidden states tapped from five evenly spaced layers of the target; its drafter-unique weights add roughly **0.5 GB at serving precision** (embeddings and output head are shared with the resident target). It follows the DSpark recipe with a diffusion-flavored block-denoising objective, survival-probability-weighted distillation, per-source-normalized hidden-state taps, and a draft block size chosen from a measured verify-cost model of the serving stack. The drafter ships 4-bit quantized — the ~1.79 GB Q4_1 pack is the default; it drafts faster than the bf16 reference at essentially unchanged draft quality, and because verification preserves the target distribution exactly, drafter precision affects only speed, never output quality. On the CUDA serving path the drafter is a measured net win — an accepted length of τ ≈ 3.6 at draft depth k = 4 turns into a **1.37x** end-to-end decode speedup on H100 (104.8 → 143.8 tok/s). On Apple Silicon the batch-1 verification pass does not yet amortize, so the drafter layer is not enabled by default on-device. Evaluated with EvalScope + vLLM on NVIDIA H100 under identical infrastructure, decoding, and scoring, in **thinking mode** — where the model's full reasoning is exercised and the sub-4-bit collapse of conventional methods is most visible. 15 benchmarks across six skill categories. For cross-family context the table also includes Gemma-4-31B, a model of the same capability tier, with its conventional low-bit builds — the collapse below 4 bits is a property of the methods, not of one base model. Bit-widths are true averages; "vs FP16" is relative to the Qwen3.6-27B FP16 reference. The aggregate gap also understates *how* the conventional builds fail: their degradation is selective, concentrated on the benchmarks that demand sustained chains of reasoning. IQ2_XXS falls to 57.5 on AIME26 and 56.4 on LiveCodeBench while still scoring 88.93 on MMLU-Redux — which is why casual testing misses the collapse. 1-bit Bonsai holds exactly these benchmarks, keeping AIME above 87 at a third of IQ2_XXS's footprint. The reasoning backbone comes through intact: math stays at 91.66 — within four points of full precision — and coding at 81.88, the behaviors that conventional sub-4-bit representations lose first. The 1-bit model trades part of the ternary model's margin on the most demanding categories for the smaller footprint that puts a 27B model on a phone. Intelligence density captures the ratio of a model's capability to its deployed size: D = -log2(1 - score/100) / size_GB 1-bit Bonsai 27B delivers roughly **2.7x** the density of the densest conventional build (IQ2_XXS at 0.199) and over **10x** FP16 — no conventional build of Qwen3.6-27B or Gemma-4-31B exceeds 0.2. Each stored gigabyte is translated into far more usable intelligen