Irfanuruchi/Nanbeige4.2-3B-CE-v1.0-MLX-6bit

Nanbeige4.2-3B-CE v1.0 — MLX 6-bit

Native MLX 6-bit release 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

Quantization

• —Format: MLX
• —Quantization: 6-bit affine
• —Group size: 64
• —Effective quantization reported by MLX-LM: 6.500 bits per weight
• —Approximate release size: 3.2 GB

This is the higher-precision MLX quantized option relative to the 4-bit release; no separate factual-quality benchmark claim is made.

The converted config preserves the native Nanbeige architecture, including:

• —model_type = nanbeige
• —num_loops = 2
• —skip_loop_final_norm = false
• —22 hidden layers
• —hidden size 3072
• —intermediate size 10752
• —48 attention heads
• —8 KV heads
• —configured context length 262144
• —vocabulary size 166144

Conversion provenance

Converted from the frozen canonical v1.0 source:

• —Source repo: Irfanuruchi/Nanbeige4.2-3B-CE-v1.0
• —Source revision: v1.0
• —Source commit: 171d26c743c6f1f5e3fd49f88c0454448a28c39e
• —MLX-LM commit: 6d21ce4b065a2e163fa6de76a9936c61aeb5784a
• —MLX-LM version: 0.32.0
• —MLX version: 0.32.2
• —Python: 3.12.14
• —Validation platform: Apple Silicon arm64

The original Transformers auto_map metadata and copied Nanbeige Python implementation files were removed after conversion because this release uses MLX-LM's native mlx_lm.models.nanbeige implementation. The MLX model weights were verified unchanged by that cleanup.

Runtime validation

Validated locally on Apple Silicon with native MLX-LM loading:

• —Model class: mlx_lm.models.nanbeige.Model
• —Inner model class: mlx_lm.models.nanbeige.NanbeigeModel
• —num_loops = 2
• —skip_loop_final_norm = false
• —Tokenizer: LlamaTokenizer

A deterministic runtime smoke test with thinking disabled generated:

Docker 127.0.0.11 is the embedded DNS resolver for user-defined Docker networks.

Observed on the validation machine:

• —Generation: about 23.8 tokens/s
• —Peak memory: about 3.6 GB

These numbers are validation observations from one Apple Silicon system, not general performance guarantees.

Current Transformers may emit a warning that it is loading a nanbeige config into an unregistered Transformers model type. In the validated MLX-LM stack this is a tokenizer/config warning; MLX-LM still selects and runs its native Nanbeige implementation.

Usage

Install a version of MLX-LM with native Nanbeige support, then run:

mlx_lm.generate \
  --model Irfanuruchi/Nanbeige4.2-3B-CE-v1.0-MLX-6bit \
  --prompt "What is Docker 127.0.0.11 on a user-defined network?" \
  --max-tokens 96 \
  --temp 0 \
  --chat-template-config '{"enable_thinking": false}'

SHA-256

See SHA256SUMS for hashes of the complete release package.

Frozen MLX weight SHA-256:

a413867f7f62e589f153134326b272c4b92931b6758f559bf746c269d7b82124

Validation and limitations

This is a deployment conversion of the frozen Nanbeige4.2-3B-CE v1.0 model. Quantization does not create a new training checkpoint or a new factual-quality release gate.

The source v1.0 checkpoint 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.