araysuter/Ripple-65B-Preview-MLX-mixed-2bit

Ripple-65B Preview — MLX mixed 2-bit

Reduced-fidelity variant. A smoke probe produced an incorrect is_prime function; prefer mixed-4bit when memory permits.

Converted independently from the original instruction-tuned BF16 checkpoint, revision 4f308b9954b73cd200812095916c7baf68cf27f4. No existing quantized model was requantized. The pretraining-only Ripple Base model was not used.

Stored component	Precision
MLP matrices	2-bit, groups of 32
Other attention projections	4-bit, groups of 64
Attention sigmoid gates, embedding and output head	8-bit, groups of 64
Norms and value-residual scalars	Original BF16

Weights occupy 29.62 GB / 27.58 GiB. The actual average is 3.657 bits per parameter, including affine scales, biases and protected tensors. The name refers to the large MLP weights, not every parameter. The encoder selected for this profile was mse. MSE optimization uses per-group clipping search and binary centroid refinement with an affine error fallback; it is not activation-calibrated GPTQ/AWQ.

Run with MLX

This release includes a custom Ripple adapter using MLX-LM's model_file support. Review modeling_ricursive_mlx.py; loading it requires the explicit trust option. Tested with MLX 0.32.4 and MLX-LM 0.32.2.

pip install mlx==0.32.4 mlx-lm==0.32.2
python -m mlx_lm.generate --model araysuter/Ripple-65B-Preview-MLX-mixed-2bit --trust-remote-code \
  --prompt 'Explain binary search in two sentences.' --max-tokens 512

The adapter preserves PeriNorm, Q/K norms, embedding scaling, local RoPE, global NoPE, value-residual blending and clipped SwiGLU. Local KV is bounded. The reference's FP32 activation computation and BF16 cast order are preserved.

Measured fidelity and runtime scope

Diagnostic	Result
Synthetic teacher-forced positions	1024
Top-token agreement with streamed original BF16	76.27%
Mean KL from original BF16	0.342162 nats
Synthetic next-token loss, BF16 / quant	2.1479 / 2.3985 nats
Uncached 2K prompt processing, one user	306.5 tokens/s
Decode on that short retrieval continuation	17.0 tokens/s (9 generated tokens)

These are small synthetic diagnostics, not GPQA, Terminal-Bench, general task accuracy, or the publisher's BF16 benchmark scores. The original weights were evaluated layer by layer through the checked MLX reference adapter; this is not a whole-model PyTorch oracle. Timing used plain MLX without a drafter, a resident model, fresh KV and 256-token prompt chunks. Short continuations do not establish sustained throughput. Raw numeric scope is in VALIDATION.json.

This profile did not receive a separate 128K capacity run; weight size is not a context-fit guarantee.

DFlash2 and TensorFold

The matching draft is Ripple-65B Preview DFlash2. It requires a Ripple-aware DFlash2 runtime; it is not a standalone Qwen3 language model. This release does not claim a tested DFlash2 path, distribution-preserving speculative decoding, or native TensorFold API support. Native Ripple bring-up is tracked in TensorFold #2171, with drafting planned separately. Target-only 128K results cannot be applied to a drafter plus native snapshot configuration.

See the conversion tools, source pins, adapter tests and limits. Keep the original model's Apache-2.0 license; the license is included here.