petergilani/Qwen3.8-Flash-Next-bf16-ngram

Qwen3.8-Flash-Next — bf16 n-gram table (a tier SOURCE, not a model)

Do not mlx_lm.load() this repository. This is not a runnable model. It is the one external weight tensor a paired build needs: the n-gram PLE embedding table of Qwen/Qwen3.8-Flash-Next, published on its own so disk-tiered builds (below) can stream it from NVMe instead of carrying it.

What is in here

• The 33 upstream shard files that host the n-gram table, byte-identical to their counterparts among the 131 shards of Qwen/Qwen3.8-Flash-Next (original model-000NN-of-00131.safetensors names, original model.language_model.* keys, original bytes — nothing renamed, nothing requantized). Because they are verbatim shards, they also still contain the small interleaved non-n-gram weights that sat beside the table in those files (~1.9 GB in total).
• model.safetensors.index.json, filtered to the table only: exactly the 128 entries model.language_model.layers.1.ple.ple_embedding.ngram_embedding.shard_{0..127}.weight, total_size = 102,400,491,520 bytes. A tier consumer reads this index plus the file headers and then touches the file bytes by range only.
• LICENSE — the upstream license (Qwen Community License 1.0).

Table geometry: 128 shards × 2,500,012 rows × 160 columns, BF16 = 320,001,536 rows, 320 bytes per row, 102.400 GB of tensor data (104.30 GB of files including the verbatim interleaved remainder and headers).

How it is used

Point a tier consumer's src_dir at a local copy (or HF snapshot) of this repo while loading its paired model, petergilani/Qwen3.8-Flash-Next-MLX-6bit: that repo ships the n-gram weights intentionally absent — 102.4 GB that no 128-GB Mac can hold resident next to a ~106 GB model — and the tier serves each needed row as a single 320-byte pread straight out of these bytes, so the table is read from NVMe instead of occupying RAM.

The pairing, load order, and verification receipts are documented on the model repo.

Provenance

Derived (by file selection + a filtered index, zero data modification) from Qwen/Qwen3.8-Flash-Next, snapshot f5d08274bafd880402bd16f5e3e6c514136ec06c. License: Qwen Community License 1.0, LICENSE included verbatim.