thoddnn/Ternary-Bonsai-27B-mlx-2bit

Prism ML Website  |  Whitepaper  |  Demo & Examples  |  Discord

Ternary Bonsai 27B

Full 27B-class reasoning in ternary transformer weights — on everyday laptops

~9.4x smaller than FP16 (ideal) | 95% of FP16 intelligence retained | ~26 tok/s on an Apple M5 Pro laptop

Highlights

• ~7.2 GB deployed footprint (down from ~54 GB FP16) — full 27B-class reasoning on a standard laptop or a single GPU
• 95% of FP16 intelligence retained: 80.49 average across 15 thinking-mode benchmarks — a higher score than the conventional IQ2_XXS build (72.73) at less than two-thirds of its footprint
• Retains thinking, reasoning, and agentic behavior deep in the sub-4-bit regime, where conventional low-bit representations collapse: math within two points of full precision (93.40), coding at 85.96, agentic tool use at 74.01
• End-to-end ternary language weights across embeddings, attention projections, MLP projections, and LM head, at a true 1.71 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
• Custom 2-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 1.34x decode speedup on the CUDA serving path
• 1-bit companion: also available as 1-bit Bonsai 27B (~3.9 GB), the phone-class operating point that fits an iPhone 17 Pro Max

Resources

• Whitepaper — full methodology, benchmarks, and measurement notes
• Demo & examples — serving, benchmarking, and integrating Bonsai
• Low-bit kernels: MLX fork (Apple Silicon) · mlx-swift fork (iOS/macOS) · llama.cpp fork (CUDA)
• Discord — join the community for support, discussion, and updates

Model Overview

Item	Specification
Base model	Derived from Qwen3.6-27B, a 27B hybrid-attention causal language model (architecture unchanged)
Parameters	~27.3B ternary language weights (~24.8B backbone across 64 blocks + ~2.5B embedding/LM head) + ~0.46B vision tower (27 blocks)
Architecture	Hybrid attention (~75% linear / ~25% full attention), SwiGLU MLP, RoPE, RMSNorm
Context length	262K tokens (full-context capable on-device, enabled by the predominantly linear-attention backbone)
KV cache	Near-lossless 4-bit KV quantization; the hybrid backbone grows a full-attention cache on only 16 of 64 layers (~4.3 GB at the full 262K window)
Weight format	Ternary g128: {−1, 0, +1} weights with FP16 group-wise scaling
Low-bit coverage	Embeddings, attention projections, MLP projections, LM head
Vision tower	HQQ 4-bit; optional ~0.63 GB mmproj pack, loaded only for image input
Deployed size	~7.2 GB (5.9 GB ideal at 1.71 bits/weight; see below)
Acceleration	DSpark speculative-decoding drafter layer provided
Backends	Apple MLX (Python, Swift) and CUDA
License	Apache 2.0

Weight Representation: Ternary g128

Each weight takes a value from {−1, 0, +1}, with one shared FP16 scale factor for every group of 128 weights. A ternary value carries log₂3 ≈ 1.585 bits of information, so the effective storage cost is ~1.71 bits/weight (ternary code + 16-bit scale amortized over 128 weights) — an idealized ~9.4x reduction vs FP16.

Relative to the binary format, the extra zero state gives a more expressive weight alphabet and recovers more of the full-precision model's behavior, which makes ternary the quality-oriented operating point of the Bonsai 27B family.

Memory Requirement

Format	True bits/weight	Ideal size	Deployed size	Reduction (ideal)
FP16 (baseline)	16.0	~54 GB	—	1.0x
Ternary g128	1.71	5.9 GB	~7.2 GB	~9.4x

Today's kernels store each ternary value in a 2-bit slot (2.125 bits/weight deployed), so the deployed footprint sits above the representation's information-theoretic minimum until native ternary kernels close the gap. 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.

Shipped Components

Two optional components ship alongside the language model (on-disk sizes from the GGUF packaging):

Component	Pack	Size	Residency
Language model	2-bit g128 slots	7.17 GB	resident
DSpark drafter	Q4_1 (default)	1.95 GB	optional — speculative decoding
DSpark drafter	bf16 (reference)	7.29 GB	optional
Vision tower	mmproj HQQ 4-bit (Q8_0 container)	0.63 GB	optional — multimodal input only
Vision tower	mmproj BF16 (reference)	0.93 GB	optional

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. A group-64 ternary pack (7.59 GB) is also published, matching the 64-value-group Q2_0 packing in llama.cpp — the same native g128 representation with each scale repeated per 64-value block.

MLX Packaging

The published MLX pack measures 8.49 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 2.25 bits/weight versus the deployed 2.125. This is a current MLX limitation; once MLX supports scale-only group formats, the MLX pack matches the native rate.

Peak Memory at Context

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):

Build	Weights	4K ctx	10K ctx	100K ctx
Ternary Bonsai (llama.cpp Q2_0)	7.15	8.4	8.7	14.7
Ternary Bonsai (MLX 2-bit)	7.57	9.2	9.6	15.5
Qwen3.6-27B "4-bit" (Q4_K_XL)	17.6	19.2	19.6	25.6
27B 16-bit (GGUF bf16)	51.25	52.6	53.3	59.3

The ternary build holds a 100K-token context at 14.7–15.5 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 ternary build's 100K peak drops to ~10.1 GB, and the full 262K window fits in ~12.8 GB peak.

Best Practices

Generation Parameters

Parameter	Suggested
Temperature	0.7
Top-p	0.95
Top-k	20

These are the settings used for all reported benchmark results (thinking mode).

System Prompt

You can use a simple system prompt such as:

You are a helpful assistant

Quickstart

MLX (Apple Silicon — Mac)

Use the Bonsai-demo repo to run ternary Bonsai 27B on MacBook.

CUDA (NVIDIA GPUs)

CUDA inference uses the GGUF pack of the same weights: fused 2-bit GEMM kernels in our llama.cpp fork unpack the ternary codes and apply the group-wise scales inside the matrix multiply — the model is never expanded into a dense FP16 tensor in memory. See Ternary-Bonsai-27B-gguf for the full CUDA, Metal, and server quickstarts.

Deploying to a phone? The ternary build (~7.2 GB) exceeds the ~6 GB per-app iOS memory budget and is laptop/GPU-only. Use the 1-bit companion (~3.9 GB), which fits an iPhone 17 Pro Max.

Cross-Platform Throughput

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. Rows are measured with llama.cpp (Metal/CUDA, custom low-bit kernels) on the GGUF pack of the same weights.

Platform	Footprint	TG128 (tok/s)	PP512 (tok/s)
Laptop (Apple M5 Max, Metal)	7.2 GB	44.0	830
Laptop (Apple M5 Pro, Metal)	7.2 GB	26.2	393
Laptop (Apple M4 Pro, Metal)	7.2 GB	18.0	125
Single GPU (H100, CUDA)	7.2 GB	98.0	2596

On the laptop, 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 interactively on an everyday laptop. The measured decode streams ~186 GB/s of weights on the M5 Pro, confirming the memory-bandwidth-dominated profile that the low-bit representation is built to exploit. 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 ternary and binary variants converge there (98 vs 104.8 tok/s) despite their ~1.9x difference in bytes per step.

Speculative Decoding: DSpark

Ternary 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 ~1.95 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 ≈ 3.7 at draft depth k = 4 turns into a 1.34x end-to-end decode speedup on H100 (98 → 131.8 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.

Benchmarks

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.

Variant	True bpw	Footprint	Thinking avg	vs FP16
Qwen3.6-27B FP16	16.0	54 GB	85.07	100%
Qwen3.6-27B Q4_K_XL ("4-bit")	5.2	17.6 GB	84.99	99.9%
Qwen3.6-27B IQ2_XXS ("2-bit")	2.8	9.4 GB	72.73	85.5%
Gemma-4-31B FP16	16.0	61.5 GB	84.58	99.4%
Gemma-4-31B QAT ("4-bit")	6.0	23.3 GB	83.41	98.0%
Gemma-4-31B Q2_K_XL ("2-bit")	3.0	11.8 GB	73.31	86.2%
Ternary Bonsai 27B	1.71	5.9 GB	80.49	94.6%
1-bit Bonsai 27B	1.125	3.9 GB	76.11	89.5%

At 5.9 GB, Ternary Bonsai 27B outscores both sub-4-bit conventional builds by more than seven points at one-half to two-thirds of their size.

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. Ternary Bonsai holds exactly these benchmarks, keeping AIME at 87.5–90.8 and LiveCodeBench at 82.8.

By Skill Category

Category	Benchmarks	FP16	Ternary 27B
Knowledge & reasoning	MMLU-Redux, MuSR	83.15	76.96
Math	GSM8K, MATH-500, AIME25, AIME26	95.33	93.40
Coding	HumanEval+, MBPP+, LiveCodeBench	88.74	85.96
Instruction following	IFEval, IFBench	78.47	71.77
Agentic / tool calling	BFCL v3, τ²-Bench	80.00	74.01
Vision	MMMU-Pro, OCR Bench v2	72.61	65.19
Overall (15)		85.07	80.49

The reasoning backbone comes through intact: math stays within two points of full precision (93.40), coding at 85.96, and the ternary model spends its extra footprint to hold the most demanding categories — agentic tool use at 74.01 and vision at 65.19 — the behaviors that conventional sub-4-bit representations lose first.

Full Per-Benchmark Results

Expand full per-benchmark results (thinking mode)

Benchmark	FP16	Ternary 27B
MMLU-Redux	93.42	88.05
MuSR	72.88	65.87
GSM8K	95.30	96.06
MATH-500	99.40	99.20
AIME25	93.29	90.84
AIME26	93.33	87.50
HumanEval+	95.12	93.90
MBPP+	83.33	81.22
LiveCodeBench	87.77	82.75
IFEval	88.91	85.03
IFBench (prompt-loose)	68.03	58.50
BFCL v3	77.10	74.41
τ²-Bench	82.90	73.61
MMMU-Pro	79.94	68.96
OCR Bench v2	65.28	61.42
Average (15)	85.07	80.49

Intelligence Density

Intelligence density captures the ratio of a model's capability to its deployed size:

D = -log2(1 - score/100) / size_GB

Variant	Size (GB)	Benchmark avg	Intelligence Density (1/GB)
1-bit Bonsai 27B	3.9	76.11	0.530
Ternary Bonsai 27B	5.9	80.49	0.400
Qwen3.6-27B IQ2_XXS	9.4	72.73	0.199
Gemma-4-31B Q2_K_XL	11.8	73.31	0.162
Qwen3.6-27B Q4_K_XL	17.6	84.99	0.155
Gemma-4-31B QAT	23.3	83.41	0.111
Qwen3.6-27B FP16	54	85.07	0.051
Gemma-4-31B FP16	61.5	84.58	0.044

Ternary Bonsai 27B delivers 2x the density of the densest conventional build (IQ2_XXS at 0.199) and nearly 8x FP16 — no conventional build of Qwen3.6-27B or Gemma-4-31B exceeds 0.2. Each stored gigabyte is translated into far more usable intelligence.

Use Cases

• Laptop-local 27B agents: full 27B reasoning and tool use on a standard laptop at ~26 tok/s, with the 262K context available for long-document analysis, full-repository code work, and other tasks that depend on holding a large working set in context
• Privacy-sensitive and offline settings: on-device execution keeps prompts and data on the device by construction, and works with intermittent or no connectivity
• Single-GPU and commodity-GPU serving: 27B-class quality from a single consumer or entry-level datacenter GPU, with headroom for larger batches, longer contexts, or co-resident models — combined with the KV-cache quantization, high-throughput serving and long-context document analysis become practical on a single 24 GB GPU
• Quality-first low-bit deployment: when the deployment target has laptop-class memory or better, ternary is the operating point that retains the most of the full-precision model's behavior

Limitations

• The quality–footprint trade-off: the ternary model retains 94.6% of the full-precision average, and the gap is modest and predictable — the reasoning core (math, coding) stays within a few points of baseline, with the difference concentrated in the most demanding categories; deployments that need the last few points of accuracy can still reach for the full-precision model where its footprint is not a constraint
• Does not fit a phone: at ~7.2 GB the ternary build exceeds the ~6 GB per-app iOS memory budget; use the 1-bit companion for phone deployment
• Served in 2-bit slots today: the deployed footprint (~7.2 GB) sits above the representation's ~5.9 GB native target; native ternary kernels are an active engineering target and would return the remaining bandwidth and footprint advantage directly as latency and energy improvements
• Agentic coding (long-horizon, multi-file, run-test-and-repair workflows) is not yet a strong target of this release; a Bonsai 27B variant tuned for agentic coding is next on the roadmap
• KV compression headroom: this release standardizes on a 4-bit KV cache; Bonsai's tolerance to KV-cache error grows with context length, and early results show the key cache can be pushed toward the sub-2-bit regime — a path to still longer contexts within a fixed device-memory budget

Citation

If you use Ternary Bonsai 27B, please cite:

@techreport{bonsai27b,
    title   = {Bonsai 27B: Full 27B-Class Reasoning in Binary and Ternary
               Transformer Weights --- on Laptops and Phones},
    author  = {Prism ML},
    year    = {2026},
    month   = {July},
    url     = {https://prismml.com}
}

Contact

For questions, feedback, or collaboration inquiries: contact@prismml.com