Downloads · 30 days
35
40% of all-time downloads
BlivionIaG/Qwen3.8-Flash-Next-EXL3-3bpw-3inst-uncalibrated
Qwen3.8-Flash-Next-EXL3-3bpw-3inst-uncalibrated is a image-text-to-text model from BlivionIaG. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for transformers. The card lists the license as other.
Uncalibrated. WIP. No LDLQ. Weights were encoded with Viterbi only, so this is not a quality release. It exists to test the opengfx1030/vllm-rdna fork. Do not treat outputs as representative of Qwen/Qwen3.8-Flash-Next.
Downloads · 30 days
35
40% of all-time downloads
All-time downloads
88
Public
Parameters
57.2B
114 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors114 GB · 100%
How the weights are stored.
I1655.4B · 97%
From the Hugging Face model README
Uncalibrated. WIP. No LDLQ. Weights were encoded with Viterbi only, so this is not a quality release. It exists to test the opengfx1030/vllm-rdna fork. Do not treat outputs as representative of Qwen/Qwen3.8-Flash-Next.
EXL3 trellis of Qwen/Qwen3.8-Flash-Next (Qwen Team). Codebook 3inst. Input Hadamard only (hadamard: in). The part split follows turboderp's mul1 3.05bpw recipe (-b 3 experts, -hq -hb 5 on the dense group). 3inst is weaker than mul1, so that dense group is K=6 instead of K=5. index_qk_proj stays bf16.
Widths that are not a multiple of 128 are zero-padded to the next block (n-gram 160→256, vision linear_fc2 4304→4352). The patch embed is stored as the flattened matrix (1152, 1536).
| Tensors | Codebook | K |
|---|---|---|
Routed experts, including MTP (mlp.experts.*) | 3inst | 3 |
Attention q/k/v/o_proj | 3inst | 6 |
Gated DeltaNet in_proj_qkv, in_proj_z, out_proj | 3inst | 6 |
| Shared expert | 3inst | 6 |
lm_head | 3inst | 6 |
PLE key_proj, value_proj; MTP fc_hidden, fc_embedding | 3inst | 6 |
| N-gram table, 128 shards | 3inst | 6 |
Vision linears (qkv, proj, linear_fc1, linear_fc2, merger) and patch embed | 3inst | 6 |
Embeddings (embed_tokens, pos_embed), norms, conv1d, A_log, dt_bias, in_proj_a, in_proj_b, hyper-connection, MoE router (.mlp.gate.), index_qk_proj, biases | bf16 | — |
3inst codebook and the Hadamard rotation used by that project. The bit allocation above is that project's mul1 3.05bpw split, shifted up one K on the dense group.