EngineeringWithJay/Swift-Bonsai-2-27B-MTPLX

Swift Bonsai 2 27B — MTPLX community build

Swift Bonsai 2 is UkisAI's reasoning-efficient derivative of Prism ML's Ternary Bonsai 2 27B. This pack repackages the Swift trunk to serve locally on Apple Silicon with speculative decoding via MTPLX.

This is a local community build, not an official release from ukisai, Prism ML, Alibaba Cloud, or Youssofal Altoukhi (MTPLX is Apache-2.0 and unaffiliated). For the text-only build see EngineeringWithJay/Swift-Bonsai-2-MLX.

What is inside

• mtplx-pack/: Swift DPO trunk in Prism's ternary "prism_hadamard_qwen35" format, plus the Qwen3.8-27B int4 g64 MTP sidecar (mtp.safetensors, unchanged from upstream), plus the base Prism-ML vision tower grafted in unchanged to satisfy MTPLX's runtime contract (ukisai released Swift as text-only; the grafted tower is NOT Swift-aligned, and vision outputs are unproven).
• mtp.safetensors blob SHA256 f803737b0e64bcce54fcfa22f2c7ee342ad37192b8e2a3299fc940387a828c9d.

Architecture

Field	Value
Backbone	Qwen3.8-27B (GQA, head_dim 256, 24 Q heads, 4 KV heads)
Trunk	Ternary weights (one scale per signed 128-block), hadamard-signed 2-bit
Attention	Hybrid Gated DeltaNet linear attention + full attention every 4th layer
Hidden	5,120
Layers	64
Context	262,144 native; engine window when served: 8,192
MTP	Qwen3.8-27B int4 g64 MTP head (depth 1 default)

Performance

Swift's own benchmark suite is upstream; this pack does not re-measure. From ukisai's published results, vs base Ternary Bonsai 2: GPQA-Diamond 84.34% (vs 84.18%), C-Eval 81.47% (vs 81.37%), IFBench 82.60% (vs 82.13%), AIME 2025 93.33% (vs 92.00%). See ukisai's model card for methodology.

MTP speculative decoding keeps the Swift trunk distribution exact at any temperature (accept/reject with residual sampling), but the MTP head is not Swift-trained; expect lower acceptance than on the base pack.

How to use

# pip install mtplx
pip install mtplx

# Serve the M5-appropriate Swift pack (8K window, MTP depth 1)
mtplx serve --model ./mtplx-pack --port 8001 \
  --reasoning-effort medium --reasoning off --allow-swap

# Endpoints
#   OpenAI:    http://127.0.0.1:8001/v1/chat/completions
#   Anthropic: http://127.0.0.1:8001/v1/messages
#   Health:    http://127.0.0.1:8001/health

Sampling (ukisai's Swift settings): temperature 1.0, top-p 0.95, top-k 20, min-p 0.0. Do not add repetition penalties.

Memory

~8.2 GB wired; 18 GB+ unified memory recommended for agent clients. On 16 GB unified memory this fits only with --allow-swap and severe swap pressure; not viable for agent turn work.

Limitations

• Experimental research release. Swift is uneven on instruction following, tool use, and open-ended coding — validate for your workload.
• Grafted vision tower is byte-identical to Prism's, not DPO-adapted to Swift.
• MTP head is off-policy (Swift DPO shift is invisible to the head); exact-sampling holds, acceptance rate regresses.
• Exact numbers for Swift-on-MTPLX have not been published; do not quote the base pack numbers as Swift's.

License & attribution

Apache-2.0. Created using Bonsai by Prism ML. Swift weights by UkisAI. Qwen3.8-27B MTP by Alibaba Cloud / Qwen team (Apache-2.0). MTPLX by Youssofal Altoukhi (Apache-2.0); if you build a product on this pack, MTPLX's terms require an "Powered by MTPLX" attribution. See NOTICE.txt and LICENSE.

Citation

@misc{swift-bonsai-2-mtplx,
  title  = {Swift Bonsai 2 27B — MTPLX community build},
  author = {EngineeringWithJay},
  year   = {2026},
  url    = {https://huggingface.co/EngineeringWithJay/Swift-Bonsai-2-MTPLX}
}