64K draft vocabularies for Talus
Four draft-vocabulary files for speculative decoding with Talus (and
llama.cpp's --spec-draft-vocab). Each file is a set of 65,536 token ids of one model's tokenizer. A drafter with a
draft vocabulary computes its output layer for these ids only, which makes each draft step cheaper. The target model
still verifies every draft against its full vocabulary, so the output does not change: a draft vocabulary only
affects how many drafts are accepted.
| File | Tokenizer of | Used by Talus for | SHA-256 |
|---|---|---|---|
qwen-dv-65536.bin |
Qwen3.8-27B | the MTP and DFlash2 drafters | a1a7c29a4262fa8658c91f2855c99d2feef443dd01f04a56aca6d8b484167457 |
glm-dv-65536.bin |
GLM-5.3-Flash | the NextN drafter | 55d63bbd101508ba4a4ec6ded0736c151dd538b52345420bd33260e45699c4df |
mimo-dv-65536.bin |
MiMo-V2.6-Flash-MOPD | the DFlash drafter | 9eead88fb62cc1a95be16abcd33b90e61840f188df850be6f777731a018d34a3 |
laguna-dv-65536.bin |
Laguna-S-2.1 | (not used by Talus's prompt-lookup drafting; for DFlash drafters in other engines) | cd718f7de9b92ed101f8b4c6bd3754e59ca2d5a59705fd6fc0b0b82e2498bceb |
Every file is listed in SHA256SUMS.
Format
Little-endian int32 token ids, sorted ascending, 65,536 of them (256 KiB per file): llama.cpp's draft-vocabulary format. Every special (non-normal) token of the tokenizer is always included.
Use it with Talus
Pass the file of the model you serve; the flags are Talus's shipped defaults, from docs/MODELS.md:
# Qwen3.8-27B: ... --spec-draft-vocab /models/qwen-dv-65536.bin
# GLM-5.3-Flash: ... --draft-vocab /models/glm-dv-65536.bin
# MiMo-V2.6-Flash: ... --spec-draft-vocab /models/mimo-dv-65536.bin
How they were made
recipe/make_vocab.py reads the tokenizer from the model's GGUF and ranks the normal tokens by weighted frequency
over four tokenized corpora: model-written agentic text (weight 0.4), documentation and prose (0.25), source code
(0.2) and Python (0.15). It keeps the special tokens plus the top-ranked tokens up to 65,536 ids. The corpora were
split 80 / 20 to check coverage on held-out text.
The corpora are private and are not published. The files hold only a sorted set of token ids; they contain no text and no frequencies.
Licenses
Each file is derived from a model's tokenizer and is released under that model's license:
qwen-dv-65536.bin: Apache-2.0 (Qwen/Qwen3.8-27B).glm-dv-65536.bin: MIT (zai-org/GLM-5.3-Flash, Copyright (c) 2026 Z.AI Co., Ltd).mimo-dv-65536.bin: MIT (XiaomiMiMo/MiMo-V2.6-Flash-MOPD, Copyright (c) 2026 Xiaomi Corporation; the license is the MIT declaration on the model card).laguna-dv-65536.bin: OpenMDW-1.1 (poolside/Laguna-S-2.1); a copy of the agreement is inLICENSE-OpenMDW-1.1.md.recipe/make_vocab.py: MIT, like Talus.
The license texts are in licenses/.