markzen/Swift-Qwen3.8-27B-NInfer-v3

Swift-Qwen3.8-27B NInfer v3

v3 container upgrade of knoopx/Swift-Qwen3.8-27B-NInfer, which is stuck on the NInfer v2 format. No weights were converted, requantized, or touched: the v2 payload is byte-identical in this file, only the directory was rewritten for v3.

How it was made

The stock tools/upgrade_ninfer_v2_to_v3.py rejects this artifact because it only accepts official (model_id, weights_id) pairs and exact object counts. The graft carries 1307 objects under ("qwen3.8-27b", "nvfp4"), which the tool's table does not list. The fix is one lookup entry:

stock.KNOWN_COUNTS[("qwen3.8-27b", "nvfp4")] = (1307,)

Everything else the tool handles. Output directory, part ranges and component configs match the official neroued/Qwen3.8-27B-nvfp4-NInfer v3 layout, minus DFlash2 which this model does not have.

File

Field	Value
Filename	Swift-Qwen3.8-27b-nvfp4-w8g32-q4g64-q5g64-q6g64-bf16.ninfer
Size	18,324,343,792 bytes (17.07 GiB)
SHA-256	84da52a62ef76ef15ef5183c34a1a2cf4d9a9c0481c04b17ed6f6ef6e9f9cd01
Container version	3
Objects	1307 (1301 tensors, 6 resources)
Bindings / uses	1422 / 785
Components	text, vision, mtp

Verify with:

printf '%s  %s\n' \
  '84da52a62ef76ef15ef5183c34a1a2cf4d9a9c0481c04b17ed6f6ef6e9f9cd01' \
  'Swift-Qwen3.8-27b-nvfp4-w8g32-q4g64-q5g64-q6g64-bf16.ninfer' | sha256sum --check

Notes

• Chat template replaced. The upgrade installs ninfer master's qwen3_8.jinja (thinking on, effort xhigh), replacing the Swift template the knoopx artifact shipped. Override at startup with --chat-template FILE if you want the original back.
• No DFlash2. MTP (--spec mtp --draft-tokens 3) is the only speculative backend here.
• Not numerically validated. Structural validation only: framing, directory, binding ranges, and payload byte-identity against the v2 source. No perplexity or KL comparison was run.
• Requires an RTX 5090 and NInfer built from current master.

Source

Original artifact and graft provenance: knoopx/Swift-Qwen3.8-27B-NInfer. Engine: Neroued/ninfer.