Qwen3.8-Flash-Next draft layer and image projector for Lamina
The multi-token-prediction (MTP) layer of Qwen3.8-Flash-Next in the runtime format of Lamina, which drafts tokens with it for speculative decoding. A draft is only a proposal: the model checks every one, so the output is the model's own. The model's image projector in Q8_0, for Lamina's image encoder, is here too (below).
| File | Bytes | SHA-256 |
|---|---|---|
Lamina-Qwen3.8-Flash-Next-MTP.gguf |
824,052,896 | eeaa94d4cd1e51f5b548aacbb72f8a0f6989cc11c62923b5cece11b33712c7db |
Put it in the model's folder: lamina-server takes it from beside the model (or beside the server) when its config names
no draft layer; elsewhere the config's args name it, "--mtp", "<folder>/Lamina-Qwen3.8-Flash-Next-MTP.gguf".
Lamina's download-draft-layer.bat / download-draft-layer.sh download it and check it, or:
hf download eddoursul/Lamina-Qwen3.8-Flash-Next-MTP Lamina-Qwen3.8-Flash-Next-MTP.gguf --local-dir <the model's folder>
The file is a GGUF of the layer alone (general.architecture qwen4exp-mtp, lamina.mtp.version 1): no model runs
from it, and it works with every quant of Qwen3.8-Flash-Next that Lamina reads.
mlp.experts.gate_up_proj,mlp.experts.down_proj: the layer's 512 routed experts from Qwen's BF16 checkpoint in Q2_0 (64-value blocks, each block's scale chosen to minimize the squared error over the levels {-1, 0, 1, 2}).- The other 29 tensors: the rest of the layer (attention, projections, hyper-connections, router, shared expert, norms), the large projections in Q8_0, the hyper-connection and router weights in BF16, the norms in F32. Its weights start from the self-distilled ones of basalt and are trained further on Lamina's own output with ISTA-DASLab's GSQ-RCO IQ3_S and unsloth's UD-Q4_K_XL (987 requests, 2.4M positions), the target each draft's chance to be accepted by the model's greedy or sampled choice.
draft_vocab: the 40,525 token ids (English and code) the draft head proposes from. Lamina'stools/mtp_rt.py --gguf <this file> --draft-vocab <ids> --out <file>writes the layer over another vocabulary.
On held-out text the drafts carry 2.489 (IQ3_S) / 2.433 (UD-Q4_K_XL) tokens a verify pass, against 2.402 / 2.344 with basalt's dense part; on one PC (RTX 3090 and RTX 5070 Ti) Lamina writes thinking text 1.3% / 4.5% faster with it.
The earlier layout
Lamina also reads the same layer as a folder of these four files (--mtp <folder>), the only form its builds before
2026-11-15 read: the same bytes, the experts in Lamina's own blob layout (gate and up rows interleaved, the Q2_0 codes
and scales in planes).
| File | Bytes | SHA-256 |
|---|---|---|
experts.bin |
707,788,800 | 09398406be61f1f54c93861f449e48b8df0bfccbc9ec9b2b7636775a6ea9244f |
dense.bin |
116,099,072 | a46561c5d3e76e78eebb9b54bbf4ae87ba1f99a62cad85562cbb516141cd5e47 |
dense.txt |
1,880 | 8773c81ebb0986e37fe94a8a9933e87be48b1fabc6889a0106bcdab179d1c2ac |
draft_vocab.bin |
162,100 | 369151522226a5edaa5f12cfd1e2ae7db8f4fbdbd222f3dcf327dced9597fb25 |
The image projector
| File | Bytes | SHA-256 |
|---|---|---|
mmproj-Qwen3.8-Flash-Next-Q8_0.gguf |
616,703,328 | 2486a2e7bd834a0494fd1f5b324f66b78f7836cb0c6644b090be7ed1c58527c3 |
Qwen3.8-Flash-Next's vision projector for llama.cpp's mtmd, which Lamina's image encoder (lamina-vision) runs:
ISTA-DASLab's BF16 file
(mmproj-Qwen3.8-Flash-Next-BF16.gguf, SHA-256 b1a82259702816a5330d7bd7607cd9676b11780e79ff7348c21103ff3ce49bd0)
converted by Lamina's tools/mmproj_q8.py with the gguf-py of llama.cpp 3cf0325: each BF16 matrix whose rows are whole
32-value blocks in Q8_0, the other BF16 tensors in F16, the F32 ones as they are, as llama-quantize converts it (its
tensors byte for byte; its embeddings of a picture identical). On the CPU it encodes a picture ~20% faster than the
BF16 file.
License
Derived from Qwen3.8-Flash-Next's weights: the Qwen Community License 1.0 applies.