f2llm-v2-0.6b GGUF

GGUF format of codefuse-ai/F2LLM-v2-0.6B for use with CrispEmbed.

F2LLM Embedding v2 0.6B. Qwen3-based, strong multilingual performance.

Files

File Quantization Size
f2llm-v2-0.6b-q4_k.gguf Q4_K 400 MB
f2llm-v2-0.6b-q5_k.gguf Q5_K 452 MB
f2llm-v2-0.6b-q8_0.gguf Q8_0 610 MB
f2llm-v2-0.6b.gguf F32 2279 MB

Parity vs HuggingFace reference

Cosine similarity vs the upstream sentence-transformers reference on a fixed test set (text):

Quant Text
q8_0 0.9941
q5_k 0.9031
q4_k 0.6999

Note: below the 0.99 retrieval-quality bar β€” text: q5_k (0.903), q4_k (0.700). Embeddings are still functionally usable (>0.9 = directionally correct for similarity ranking) but expect small differences in nearest-neighbor results vs the upstream f32 reference.

Quick Start

# Download
huggingface-cli download cstr/f2llm-v2-0.6b-GGUF f2llm-v2-0.6b-q4_k.gguf --local-dir .

# Run with CrispEmbed
./crispembed -m f2llm-v2-0.6b-q4_k.gguf "Hello world"

# Or with auto-download
./crispembed -m f2llm-v2-0.6b "Hello world"

Model Details

Property Value
Architecture Qwen3
Parameters 600M
Embedding Dimension 1024
Layers 28
Pooling last-token
Tokenizer GPT-2 BPE
Base Model codefuse-ai/F2LLM-v2-0.6B

Verification

Verified bit-identical to HuggingFace sentence-transformers (cosine similarity >= 0.999 on test texts).

Usage with CrispEmbed

CrispEmbed is a lightweight C/C++ text embedding inference engine using ggml. No Python runtime, no ONNX. Supports BERT, XLM-R, Qwen3, and Gemma3 architectures.

# Build CrispEmbed
git clone https://github.com/CrispStrobe/CrispEmbed
cd CrispEmbed
cmake -S . -B build && cmake --build build -j

# Encode
./build/crispembed -m f2llm-v2-0.6b-q4_k.gguf "query text"

# Server mode
./build/crispembed-server -m f2llm-v2-0.6b-q4_k.gguf --port 8080
curl -X POST http://localhost:8080/v1/embeddings \
    -d '{"input": ["Hello world"], "model": "f2llm-v2-0.6b"}'

Credits

Provenance and EU AI Act Art. 53 note

  • Upstream model: codefuse-ai/F2LLM-v2-0.6B β€” published by codefuse-ai.
  • Upstream licence: apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
  • What was done here: format conversion and/or quantisation only (GGUF/GGML). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
  • Training data: documented β€” where it is documented at all β€” by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
  • Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.

Revision 2026-11-02: f2llm-v2-0.6b-q4_k-mixed.gguf supersedes the Q4_K files

The CUDA/CPU audit measured f2llm-v2-0.6b-q4_k-imatrix.gguf at cosine 0.78 against the F32 file, below the Q4_K acceptance bar of 0.94 (-q4_k-imatrix-c2 0.76, plain -q4_k 0.59). They remain here unchanged for reproducibility but are superseded. The replacement keeps Q4_K+imatrix (134-text calibration corpus) for most matrices, with ffn_down at Q6_K, attn_v and the last three blocks (25-27) at Q8_0, and token embeddings at Q8_0. On a 78-text panel (13 audit texts plus 65 held-out German/English/code/query texts), minimum cosine against codefuse-ai/F2LLM-v2-0.6B (sentence-transformers, FP32) is about 0.96, versus 0.77 for the superseded file. The imatrix used is f2llm-v2-0.6b-q4_k-mixed.imatrix. Evidence: CrispEmbed tests/results/below-bar/.

Downloads last month
577
GGUF
Model size
0.6B params
Architecture
qwen3
Hardware compatibility
Log In to add your hardware

4-bit

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for cstr/f2llm-v2-0.6b-GGUF

Finetuned
Qwen/Qwen3-0.6B
Quantized
(9)
this model