xunlinkx/Qwopus3.8-27B-Coder-Flash-Text-oQ4e-fp16-mtp

Qwopus3.8-27B-Coder-Flash-Text — MLX oQ4e-fp16-mtp (Pure Text + MTP)

This is an Apple MLX enhanced oQ4 (oQ4e-fp16-mtp) quantization of Jackrong/Qwopus3.8-27B-Coder-Flash, prepared by xunlinkx using the oMLX quantization pipeline.

This Text-Only edition cleanly omits the visual transformer weights and multimodal processor configs to achieve the lowest possible memory footprint and highest generation speed on 18 GB, 24 GB, and 32 GB Apple Silicon Macs. Tailored specifically for headless coding agents, terminal CLI harnesses (Hermes, OpenCode), and multi-turn software development workflows.

---

Model Details

• Base Model: Jackrong/Qwopus3.8-27B-Coder-Flash
• Architecture: qwen3_5_text / Pure Text LLM
• Parameters: ~27B dense foundation model
• Modality: Pure Text-Only. All 27 ViT visual blocks (model.visual.*), image projection tokens, and video processor configs are stripped, saving ~2.2 GB of unified memory overhead.
• Estimated Weight Footprint: ~13.8 GB on disk/RAM.
• Quantization: 4-bit enhanced mixed precision (oQ4e) with affine quantization, group size 64, and strict importance-matrix (imatrix) calibration (128 samples, 512 sequence length).
• Target Dtype (fp16): Non-quantized weights, scales, biases, norms, and embeddings are preserved in FP16 (float16), providing peak hardware execution speed on Apple Silicon M1/M2 chips while running at maximum bandwidth on M3/M4/M5 hardware.
• Speculative Decoding (MTP): Preserves the official multi-token prediction draft heads (mtp.*) in FP16 precision. When enabled in supported runtimes (oMLX, llama.cpp MTP draft), speculative decoding achieves ~1.4x–1.7x faster token throughput on coding and agentic reasoning tasks.
• License: Apache-2.0 (inherited from the upstream base model).

---

Key Configurations

Modality & Tensors

• Text Optimized: Eliminates image cache allocations and vision encoder overhead for faster cold-starts and maximum context-window headroom on 24 GB Macs.
• Multi-Token Prediction: Native auxiliary draft heads preserved in mtp.* with mtp_num_hidden_layers: 1 in config.json.
• Layer Sensitivities: Critical attention projections (v_proj, q_proj) and MLP down-projections on sensitive layers are boosted to 5-bit/6-bit to protect reasoning depth and indentation adherence.

---

Usage

With oMLX Server

Add to your ~/.omlx/model_settings.json:

{
  "Qwopus3.8-27B-Coder-Flash-Text-oQ4e-fp16-mtp": {
    "mtp_enabled": true,
    "mtp_num_draft_tokens": 3,
    "temperature": 0.2,
    "top_p": 0.95
  }
}

Start serving:

omlx serve Jackrong/Qwopus3.8-27B-Coder-Flash-Text-oQ4e-fp16-mtp --port 8005

Direct Generation with MLX-LM

mlx_lm.generate \
  --model Jackrong/Qwopus3.8-27B-Coder-Flash-Text-oQ4e-fp16-mtp \
  --prompt "<|im_start|>user\nWrite a thread-safe LRU cache in Python with O(1) ops.<|im_end|>\n<|im_start|>assistant\n" \
  --max-tokens 512 \
  --temp 0.2

---

Attribution

Base checkpoint and coder-flash fine-tuning by Jackrong. Quantization, text optimization, and MLX packaging by xunlinkx using the oMLX cluster pipeline.