Irfanuruchi/Nanbeige4.2-3B-CE-v1.0-GGUF

Nanbeige4.2-3B-CE v1.0 — GGUF

GGUF builds of Nanbeige4.2-3B-CE v1.0, a Computer Engineering and systems-focused fine-tune of Nanbeige/Nanbeige4.2-3B.

Canonical merged BF16 release:

Irfanuruchi/Nanbeige4.2-3B-CE-v1.0

Available GGUF files

Q4_K_M is the recommended compact general-use build.

Q6_K provides a higher-quality size/performance balance.

Q8_0 is a high-fidelity quantized build.

BF16 is the GGUF reference representation.

Conversion

Converted from the frozen local merged BF16 v1.0 release using llama.cpp commit:

73a43d1f69345aee8bb186ef4b3172cef892f2e5

The converter recognized the model as native GGUF architecture:

nanbeige

Important architecture metadata preserved:

• —22 blocks
• —3072 embedding dimension
• —10752 feed-forward dimension
• —48 attention heads
• —8 KV heads
• —262144 configured context length
• —num_loops = 2
• —skip_loop_final_norm = false

All released GGUF files were successfully generated with llama.cpp and locally runtime-smoke-tested.

SHA-256

See SHA256SUMS for the exact hashes of every GGUF artifact.

Usage

Example with llama.cpp:

llama-cli \
  -m Nanbeige4.2-3B-CE-v1.0-Q4_K_M.gguf

The model contains its tokenizer and chat template in GGUF metadata.

Validation and limitations

These GGUF builds are format/runtime conversions of the frozen v1.0 model. Quantization does not constitute a new training checkpoint or a new factual-quality release gate.

Nanbeige4.2-3B-CE v1.0 is the best validated release checkpoint selected during development, but it is not claimed to be perfect. Some known precision/factual weaknesses remain in difficult systems questions.

Refer to the canonical BF16 model card for the primary validation notes and limitations.

License

Apache-2.0. See LICENSE.

Base model: Nanbeige/Nanbeige4.2-3B.