Qwen3.8-Flash-Next UD-IQ3_XXS + Q8_0 n-gram, as a Basalt pack
A Basalt .basalt model file (format version 2, needs
Basalt 0.2.0 or later) built from Unsloth's UD-IQ3_XXS GGUF of Qwen3.8-Flash-Next.
This is the stock model. It is not abliterated.
What is in the file
| Part | Source |
|---|---|
| Routed experts and dense tensors | unsloth/Qwen3.8-Flash-Next-GGUF UD-IQ3_XXS shards 1-3, copied byte for byte |
| Hyper-connection projections | Stored as Q8_0 in Unsloth's GGUF; the engine reads BF16, so they were converted with convert.py --compat-bf16 (recorded in meta) |
Per-layer n-gram table (per_layer_token_embd.weight) |
Q8_0, the same table the official Basalt packs use (Unsloth's shard 4 stores it as IQ4_NL) |
| MTP drafter, vision projector | Copied byte for byte from the official Basalt pack lineage via --sections-from |
| Tokenizer, chat template | From the Unsloth GGUF |
Expert types in this quant: gate/up IQ2_S, IQ3_XXS, IQ4_XS; down IQ4_NL and Q2_0. Size 105.4 GiB (46.8 GiB of experts,
54.9 GiB in the gguf section, most of it the n-gram table).
How it was built
python3 tools/convert/convert.py \
--gguf Qwen3.8-Flash-Next-UD-IQ3_XXS-00001-of-00004.gguf \
--sections-from BASE.basalt --compat-bf16 \
--out Qwen3.8-FN-UD-IQ3_XXS-q8ple.basalt
Shard 4 of the Unsloth set (the n-gram table) was replaced by a GGUF holding the Q8_0 table before conversion; the
table bytes were copied unchanged from an existing Q8_0 Basalt pack. Converted at Basalt 1b3f6964 (0.2.1).
Measured here
3x RTX 5060 Ti 16 GB, 62 GiB RAM, Basalt 0.2.1, --spec 8, relaxed-drafting 0.2, 44 GiB RAM budget (all experts
resident, no SSD spill). Thinking off, fresh prompts.
| Workload | tok/s |
|---|---|
| ~3.5k-token prefill | 1,742 |
| ~27k-token prefill | 2,410 |
| ~168k-token prefill | 2,255 |
| Prose decode | 105-118 |
| Code decode | 159-202 |
A 15-test agent harness passed 15/15 on all of 3 runs. That is a no-regression check, not an accuracy benchmark. For quality numbers on the underlying quant see Unsloth's published comparison.
A pack you convert yourself is not the official file; the drafter here comes from the official lineage, which is why decode speed matches it.
License
Qwen community license; see LICENSE.