nightscape/Intern-S2-Mobius-4bit-mlx

Intern-S2-Mobius (4-bit MLX)

An MLX 4-bit (affine, group size 64) quantization of internlm/Intern-S2-Mobius — a 35B hybrid model: Gated-DeltaNet linear attention / full attention at interval 4, with 2560 experts in 4 globally-shared routed MoE banks. Runs on Apple Silicon.

• —Base model: internlm/Intern-S2-Mobius (Apache-2.0).
• —Quantization: 4-bit affine, group size 64; shared-expert / router gates kept at 8-bit. ~19.6 GB peak, ~4.5 bpw.
• —Architecture: interns2_mobius (text_config.model_type = interns2_mobius_text), 40 layers, head_dim 256, MoE 2560 experts / top-8, num_blocks 4. max_position_embeddings 262144.

Quick start

Requires the mlx-lm build that ships the interns2_mobius architecture (official part of mlx-lm as of the model-support PR):

pip install -U mlx-lm
mlx_lm.generate --model nightscape/Intern-S2-Mobius-4bit-mlx \
  -p "The secret to baking a good cake is" -m 1024 --trust-remote-code

--trust-remote-code is mandatory: the checkpoint bundles a custom tokenizer (tokenization_interns1.py) and model code.

Details

• —This is an MLX conversion; transformers does not yet ship interns2_mobius. The reference is the upstream repo's trust_remote_code implementation, verified by full bf16 logit diff (argmax agreement 38/39, the sole miss a bit-identical tie).
• —Capability spot-check (this 4-bit conversion): MMLU-Pro 88.3% ± 4.1 (n=60) vs upstream bf16 89.05; GSM8K 97–98% (n=100). Short-generation-budget evals truncate chain-of-thought and depress scores — an eval artifact, not a capability drop.
• —Text-only: this conversion is the language model. The upstream checkpoint is tagged image-text-to-text, but the MLX port loads the language model and generates text (no vision tower on this path).

Companion

An experimental MTP (Multi-Token-Prediction) head is published separately: nightscape/Intern-S2-Mobius-4bit-mlx-mtp. It is consumed by the omlx server's interns2_mobius MTP driver, not by stock mlx-lm.

License

Weights and code under Apache-2.0 — see LICENSE. Model by InternLM; this is a derivative conversion of their weights plus the MLX port.