robgaleano/Swift-1.5-Qwen3.8-27B-MLX-8bit

UkisAI BF16 Source  •  Companion 8-bit MTP Drafter  •  UkisAI Swift 1.5 Collection  •  GGUF  •  UkisAI Website  •  Evaluation  •  License & Attribution

Swift 1.5 Qwen3.8-27B — 8-Bit Apple Silicon MLX (+ Companion MTP Drafter)

This repository provides a full multimodal (Vision + Text) 8-bit Apple Silicon MLX quantization (group_size=64, bits=8, mode=affine, 8.627 bits per weight) of UkisAI's Swift-1.5-Qwen3.8-27b, converted using mlx-vlm 0.7.8 along with a standalone 8-bit Multi-Token Prediction (MTP) speculative decoding drafter (robgaleano/Swift-1.5-Qwen3.8-27B-MTP-MLX-8bit).

• Adapted Model (Swift 1.5 Qwen3.8-27B): Created by UkisAI from Qwen/Qwen3.8-27B (Alibaba Cloud). Uses 41.9%–58.5% fewer thinking tokens at xhigh reasoning effort while outperforming base Qwen3.8-27B on coding (81.71% vs 76.76% on LiveCodeBench v6) and agentic CLI benchmarks (72.13% vs 69.21% on Terminal-Bench 2.1).
• Why 8-Bit MLX (group_size=64)?: UkisAI's official collection includes 3-bit, 4-bit, and 5-bit MLX exports. This 8-bit (8.63 bpw) MLX conversion fills the high-precision tier for 64 GB–128 GB+ Apple Silicon Macs (M-series Max / Ultra), preserving near-BF16 reasoning stability (KLD ≈ 0.00055) and strict XML/JSON tool-calling accuracy for coding harnesses (Claude Code, Roo Code, DeepSeek Harness, OpenCode).
• 2.3×–2.5× Speculative Decoding via Companion MTP Drafter: Pairing this model with robgaleano/Swift-1.5-Qwen3.8-27B-MTP-MLX-8bit (--draft-model) boosts generation speed on Apple M5 Max (614 GB/s) from ~18.8 tok/s to 38.5–50.4 tok/s with zero quality loss.

---

Apple Silicon M5 Max (40-Core GPU, 128 GB) Benchmarks (mlx-vlm 0.7.8)

Configuration	Active GPU Memory	Decode Speed (tok/s)	Speedup	Notes
Swift-1.5-Qwen3.8-27B-MLX-8bit (Base 1 tok/step)	27.50 GiB	18.8 tok/s	1.00×	Memory-bandwidth bound (614 GB/s ÷ 29.5 GB)
+ Swift-1.5-Qwen3.8-27B-MTP-MLX-8bit (--draft-block-size 4)	28.22 GiB	38.5 tok/s avg (50.4 tok/s burst)	2.05×–2.68×	Lossless speculative decoding (xhigh effort)
Multi-Turn Warm Prompt (APC_ENABLED=1, 25K–50K ctx)	28.2–45.6 GiB	43.1–48.5 tok/s	Up to 30× faster TTFT	1.06s–2.11s warm TTFT via Automatic Prefix Caching

---

Quickstart (mlx-vlm)

pip install -U mlx-vlm

1. Start OpenAI & Anthropic Compatible Local Server (with MTP Speculative Decoding & APC Cache)

export APC_ENABLED=1
export MLX_VLM_DEFAULT_REASONING_EFFORT=xhigh

python -m mlx_vlm.server \
  --model robgaleano/Swift-1.5-Qwen3.8-27B-MLX-8bit \
  --draft-model robgaleano/Swift-1.5-Qwen3.8-27B-MTP-MLX-8bit \
  --draft-block-size 4 \
  --prefill-step-size 4096 \
  --max-tokens 32768 \
  --port 8080

2. Direct CLI Generation

python -m mlx_vlm.generate \
  --model robgaleano/Swift-1.5-Qwen3.8-27B-MLX-8bit \
  --max-tokens 4096 \
  --prompt "Implement a thread-safe LRU cache in Python with O(1) get and put."

---

Evaluation (From UkisAI's Official Swift 1.5 Release)

The external results below (from ukisai/Swift-1.5-Qwen3.8-27b) compare Qwen3.8-27B (the foundation base model) and Swift 1.5 at xhigh reasoning effort:

Benchmark	Final score		Mean tokens			Median tokens
	Qwen3.8	Swift 1.5	Qwen3.8	Swift 1.5	Reduction	Reduction
General reasoning
GPQA-Diamond	88.28%	88.59%	15,014	8,717	↓ 41.9%	↓ 58.5%
C-Eval	90.00%	90.92%	1,492	819	↓ 45.1%	↓ 16.9%
IFBench	73.53%	72.07%	8,052	4,955	↓ 38.5%	↓ 47.3%
ERQA	67.45%	65.40%	4,137	1,906	↓ 53.9%	↓ 56.2%
Mathematics
AIME 2026	98.67%	96.00%	22,014	13,203	↓ 40.0%	↓ 48.5%
HMMT November 2025	99.33%	97.33%	22,032	14,957	↓ 32.1%	↓ 47.8%
Coding
LiveCodeBench v6	76.76%	81.71%	11,184	8,448	↓ 24.5%	↓ 46.3%
Agent tasks
Terminal-Bench 2.1*	69.21%	72.13%	52,265	43,733	↓ 16.3%	↓ 0.1%

---

Related Swift 1.5 Repositories

Format / Precision	Repository	Runtime
BF16 Master (Official)	ukisai/Swift-1.5-Qwen3.8-27b	Transformers / vLLM / SGLang
MLX 8-bit Multimodal (This Repo)	robgaleano/Swift-1.5-Qwen3.8-27B-MLX-8bit	Apple MLX (mlx-vlm)
MLX 8-bit MTP Drafter (Companion)	robgaleano/Swift-1.5-Qwen3.8-27B-MTP-MLX-8bit	Apple MLX (mlx-vlm --draft-model)
MLX 5-bit (Official)	ukisai/Swift-1.5-5bit-MLX	Apple MLX
MLX 4-bit (Official)	ukisai/Swift-1.5-4bit-MLX	Apple MLX
MLX 3-bit TextOnly (Official)	ukisai/Swift-1.5-3bit-MLX-TextOnly	Apple MLX
GGUF (Official)	ukisai/Swift-1.5-Qwen3.8-27B-GGUF	llama.cpp
GSQ-RCO GGUF (Official)	ukisai/Swift-1.5-Qwen3.8-27B-GSQ-RCO-GGUF	llama.cpp

---

License and Attribution

• Adapted Model (Swift 1.5 Qwen3.8-27B): Created by UkisAI and licensed under the Swift Open License v1.0.
• Foundation Base Model (Qwen3.8-27B): Created by Alibaba Cloud (Qwen Team) and licensed under the Apache License 2.0.
• 8-Bit MLX Quantization & MTP Extraction: Converted from ukisai/Swift-1.5-Qwen3.8-27b by @robgaleano using mlx-vlm 0.7.8 (--mtp -q --q-bits 8 --q-group-size 64). See NOTICE for the full Section 4(b) modification and attribution notice.

Personal, research, educational, evaluation, and commercial use are free for individuals and organizations with gross annual revenue, including affiliates, of up to US$1,000,000 under the Swift Open License v1.0. Above that threshold, commercial use requires a separate Swift Enterprise License from UkisAI.

Citation

@misc{swift-1.5-qwen3.8-27b,
  title  = {Swift 1.5 Qwen3.8-27B},
  author = {UkisAI},
  year   = {2026},
  url    = {https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27b}
}