F2LLM-v2-330M Q8_0 GGUF

Q8_0-quantized GGUF build of codefuse-ai/F2LLM-v2-330M, distributed for use as the bundled embedding model in Spelunk.

What this is

A single GGUF file containing the F2LLM-v2-330M weights quantized from the original BF16 safetensors to Q8_0:

Projection matmul weights and the token-embedding table → Q8_0
RMSNorm weights → F32 (kept full precision; negligible size)

This is the same artifact spelunk-server would otherwise produce on-device on first run; distributing it pre-quantized cuts the first-run download from ~638 MB to ~339 MB and keeps steady-state disk use at ~339 MB.

The repository also carries the model's tokenizer.json, redistributed unmodified from the same pinned upstream revision, so a client can fetch both the weights and the tokenizer from here without contacting the upstream repo.

File	Approx. size	sha256
`f2llm-v2-330m-q8_0.gguf`	~339 MB	`2c12aad2951f1d9a3b457f890a2586d1ee19b755b377c0fb424e856e615b8f2b`
`tokenizer.json`	~8 MB	`7e295e5bb91a3d35335f92fa4294a6e4e0ab4aa586db853e14312a62135bfddc`

Provenance

Base model: codefuse-ai/F2LLM-v2-330M
Pinned source revision: 1239cdd544b24c247ed75df2ae22e5a401ac4659
Quantization: Q8_0 from the original BF16 weights (see above)
Architecture / dim: Qwen3 decoder, 896-dim embeddings

License & attribution

Licensed under the Apache License, Version 2.0, inherited from the base model. The full license text is in LICENSE and the modification notice is in NOTICE.

This is a modified copy of codefuse-ai/F2LLM-v2-330M: the weights have been quantized to Q8_0 from the original BF16 weights. The bundled tokenizer.json is redistributed unmodified. No other modifications were made.

Usage in Spelunk

spelunk-server downloads the GGUF and tokenizer.json directly from this repository (the default, or wherever SPELUNK_EMBEDDER_GGUF_REPO points). The model config is embedded in the binary, so no upstream fetch is required at runtime.