PocketAiHub/Qwopus3.6-27B-v2-MTPLX-BF16

Qwopus3.6 27B v2 MTPLX BF16

This is the full-precision BF16 MTPLX release of Jackrong/Qwopus3.6-27B-v2, pinned to source revision d0d82f4ccc9d41d4fe9595e96be4595327bb5de7. It packages the 26.896B-parameter language-model trunk in BF16 together with the model's native BF16 multi-token-prediction (MTP) sidecar for speculative decoding on Apple Silicon.

This is the BF16 master artifact used to derive smaller quantized builds. No body quantization is applied.

What MTPLX adds

The model includes mtp.safetensors and a verified MTP runtime contract. MTPLX drafts multiple future tokens with the native MTP head, then verifies them against the full model. The output distribution remains target-verified; the MTP head is an inference accelerator, not a separate replacement model.

This package requires MTPLX 2.4.0 or newer for native MTP acceleration. Stock mlx-lm does not consume the split MTP sidecar by itself.

Quick start

Install the runtime:

python3 -m pip install "mtplx==2.4.0"

Start the local OpenAI-compatible server with the verified settings:

mtplx quickstart \
  --model PocketAiHub/Qwopus3.6-27B-v2-MTPLX-BF16 \
  --download \
  --profile sustained \
  --depth 2 \
  --port 8000

PocketAI also supports this artifact through its managed MTPLX runtime. This BF16 release is very large: the repository is about 54.7 GB before runtime and KV-cache overhead. It was verified on a 128 GB Apple Silicon Mac; lower-memory configurations have not been qualified.

Verification

The release was forged and tested with MTPLX 2.4.0 on an Apple M5 Max with 128 GB unified memory. The sustained profile selected depth 2.

Decode mode	Throughput	Relative to AR	Acceptance by MTP position
Autoregressive	9.333 tok/s	1.000x	—
MTP depth 1	11.835 tok/s	1.268x	96.62%
MTP depth 2	14.569 tok/s	1.561x	94.96%, 91.60%
MTP depth 3	12.409 tok/s	1.330x	94.41%, 86.01%, 76.92%

These are single-machine forge-verification measurements, not universal performance guarantees. Throughput varies with hardware, prompt length, sampling settings, thermals, and runtime version. The complete structured evidence is in mtplx_runtime.json.

Artifact layout

• model-00001-of-00011.safetensors … model-00011-of-00011.safetensors — BF16 language-model trunk
• model.safetensors.index.json — trunk weight index
• mtp.safetensors — BF16 native MTP sidecar
• mtplx_runtime.json — MTPLX contract, forge provenance, and verification evidence
• tokenizer, template, and model configuration files

Scope and limitations

• This repository is the text-generation MTPLX artifact. It does not include the upstream vision tower/projector, so image input is not supported by this package.
• It is an experimental community model and has not received a comprehensive safety evaluation. Review outputs before using them in high-impact settings.
• Reasoning and tool-call behavior come from the upstream Qwopus model. See the upstream model card for its training description, datasets, intended use, and limitations.

License and attribution

Released under the Apache License 2.0, following the upstream model. This conversion/packaging release is maintained by PocketAiHub. MTPLX is developed by Youssof Altoukhi.