Models

The pinned Qwen3.8 Flash-Next build and draft, file verification, context limits and the separate model licences for a tokspur installation.

A file opens into a stepped set of thin plates above a processor board.

Compaibiliy

The pinned target is Qwen3.8 Flash-Next in the unpruned UD-Q4_K_XL GGUF build, with its native MTP sidecar. tokspur runs text only, keeps this model resident and serves it under the identifier tokspur within a 65,536-token context that includes the output reservation.

Pinned assets
AssetSourceFiles
Target modelunsloth/Qwen3.8-Flash-Next-GGUF, variant UD-Q4_K_XL, revision 766911a4 GGUF shards, about 111 GB
Native MTP sidecarunsloth/Qwen3.8-Flash-Next-GGUF, revision 766911aMTP/mtp-Qwen3.8-Flash-Next-shared-Q8_0.gguf, 2.79 GB

The exact file names, sizes and SHA-256 digests are in profiles/flash-next-assets.json in the checkout. The files come from the unsloth GGUF repository.

Prepare he files

  • Put the shards in a directory of their own, and put the sidecar in another directory under its MTP/ folder.
  • Use revision 766911a for every shard and the sidecar. Files on the model repository’s default branch can differ from the pins.
  • Run tokspur adopt with both directories. It checks every file's size and digest against the profile, derives tokenizer.json and chat_template.jinja from the verified GGUF metadata, checks their pinned digests and records the adoption. The derived text files live under .tokspur/state/gguf/derived-text/, so separate text downloads are unnecessary.
  • Run tokspur pack build to write the lossless GPU weight pack to .tokspur/derived/tokspur-pack-v2.asa.gguf while keeping at least 54 GB of disk space free and leaving the adopted files untouched.
  • Keep the files immutable afterwards. The server rechecks their identity at every start and refuses a changed file until you adopt again.
  • Keep the original files when trying a conversion or another numerical format.

Conex and memory

The profile sets 65,536 tokens of context, a BF16 attention cache, FP32 recurrent state and full verification of every proposed token. Larger contexts change the memory needed for model state, and the agent's messages, tool schemas and tool results all count against the limit.

Plan the memory budget

Speed mode drafer

The fastest measured speed mode uses a separate DFlash2 draft model. It only proposes tokens, and the target model checks them all before any is kept. tokspur does not download it. Fetch the pinned revision yourself and keep it in a directory of its own.

Optional speed mode asset
AssetSourceFiles
DFlash2 drafterincoai/Qwen3.8-Flash-Next-DFlash2, revision a22a174config.json and model.safetensors, 1.21 GB

The file sizes and SHA-256 digests are under legacy_dflash2_draft in profiles/flash-next-assets.json. Check them before first use. The drafter carries its own licence terms.

Model licences

The pinned target and shared MTP sidecar use Qwen Community License 1.0. Its terms require a separate licence for commercial use by a licensee or affiliate conducting a Model as a Service or AI Work Assistant business. Its internal-use exception applies only when the model, its capabilities and its outputs are unavailable to outside parties. Check the definitions and your intended activity before use.

Copies must retain the notice. Commercial products above 100 million monthly active users or US$20 million monthly revenue must display the model name prominently. The optional DFlash2 card declares Apache-2.0 and the same Qwen base, which does not establish an exception to the base conditions. Model rights remain separate from the runtime licence.