---
library_name: mlx
license: apache-2.0
base_model: Vontra/Qwen3.8-Flash-Next-MLX-oQ4-MTP
tags:
  - mlx
  - qwen
  - qwen4
  - mtplx
  - speculative-decoding
  - q4
---

# Qwen3.8 Flash-Next MLX oQ4 + MTP for MTPLX

This repository republishes the unchanged model artifacts from
[`Vontra/Qwen3.8-Flash-Next-MLX-oQ4-MTP`](https://huggingface.co/Vontra/Qwen3.8-Flash-Next-MLX-oQ4-MTP)
at source revision `43a82b3f0ff64fa417fd09ca046580f08d19b0d6`, together with
MTPLX validation metadata and guarded run receipts.

No tensor was quantized, requantized, or repacked for this upload. Target MoEs
and the embedded MTP MoEs remain resident through the normal MLX runner. Only
ngram lookups use MTPLX's bounded exact-row SSD cache. The published ngram
tensors retain the source-native affine Q4/group-32 storage.

## MTPLX runtime

Implementation and review are in
[`youssofal/MTPLX#368`](https://github.com/youssofal/MTPLX/pull/368).
The ngram cache has a configurable 10 GiB payload ceiling. Construction
measures base residency once and selects the smaller safe payload after
reserving KV/MTP, Metal working memory, safety margin, metadata, hash/route
tables, alignment, and transient buffers. LRU eviction affects residency only;
SSD rows remain authoritative. The server reserves its selected context and the full supported live prefill-chunk range before MLX load. Cache slots use a packed O(1) free chain/linked LRU, and each acquisition submits one bounded I/O task.

## Guarded validation

Validated on Apple M5 Max unified memory through the canonical
`/tmp/mtplx-gpu-exclusive.lock` wrapper:

| Workload | Result |
|---|---:|
| Prompt tokens | 16,384 |
| Generated tokens | 1,024 exactly |
| MTP accepted by depth | [241, 126, 58] |
| Prefill | 396.594 tok/s |
| Decode | 27.186 tok/s |
| Wall time | 102.297 s |
| Measured base residency | 80,435,213,384 bytes |
| Selected ngram payload | 1,677,721,600 bytes |
| Ngram cache overhead | 892,502,016 bytes |
| KV/MTP reserve | 637,046,792 bytes |
| Projected residency | 87,937,451,088 bytes |
| Peak memory | 85,222,315,432 bytes (79.37 GiB) |
| Runtime target | 88,046,829,568 bytes (82 GiB) |

The generated completion token hash is
`4d876b2e83ed4f4ab63027955b3d7b5494200d837564c7077791927d69e8695f`.
Raw smoke and 16K/1K receipts are included under `receipts/`.

## Download

```bash
mtplx pull OpensourceWTF/Qwen3.8-Flash-Next-MTPLX-oQ4-MTP
```

`mtplx pull` finalizes the manifest-backed ngram source files read-only so the retained-descriptor verifier can install the exact SSD row lane. Use the model with the Qwen4 path from MTPLX PR #368. Real MLX execution on the
validation host was performed only through the canonical exclusive GPU guard.

## Provenance

- Source repository: `Vontra/Qwen3.8-Flash-Next-MLX-oQ4-MTP`
- Source revision: `43a82b3f0ff64fa417fd09ca046580f08d19b0d6`
- MTPLX receipt commit: `446c95f651663c390097bf05e09c7a66045756df`
- Safetensor shards: 22
- Artifact bytes measured locally: 113,324,747,928
- Source files matching the pinned revision by size and Hub blob identity: 33
- Destination-only MTPLX manifest/validation/receipt files: 7

All 22 safetensor shards and the remaining source model/config files match the
pinned source revision. The only source-named metadata differences are this
MTPLX model card and Hub-generated `.gitattributes`.

See `MTPLX_VALIDATION.json` for machine-readable provenance and metrics.