Inkling-0.6B-A0.6B

This is a tiny version of thinkingmachines/Inkling created for testing and development.

Model Details

Base Model: thinkingmachines/Inkling
Architecture: inkling_mm_model (InklingForConditionalGeneration)
Total Parameters: 0.644B
Activated Parameters: 0.602B

Configuration Changes

The following parameters were reduced from the original model:

Parameter	Original	Tiny
`text_config.num_hidden_layers`	66	12
`text_config.hidden_size`	6144	1024
`text_config.intermediate_size`	24576	4096
`text_config.num_attention_heads`	64	8
`text_config.num_key_value_heads`	8	2
`text_config.swa_num_attention_heads`	64	8
`text_config.swa_num_key_value_heads`	16	4
`text_config.n_routed_experts`	256	8
`text_config.num_experts_per_tok`	6	4
`text_config.moe_intermediate_size`	3072	512
`text_config.num_mtp_layers`	8	1
`vision_config.n_layers`	4	1
`vision_config.hidden_size`	1024	256
`vision_config.decoder_dmodel`	6144	1024
`audio_config.decoder_dmodel`	6144	1024

Layer type patterns are preserved: 2 repetitions of [5× hybrid_sliding + 1× hybrid], with the first 2 MLP layers as dense and the rest as sparse (MoE).

Checkpoint Structure

Single safetensors file (model.safetensors). Key naming matches the original checkpoint format (model.llm.*, model.audio.*, model.visual.*).

Usage

from transformers.models.inkling import InklingForConditionalGeneration
from transformers import AutoTokenizer

model = InklingForConditionalGeneration.from_pretrained("inference-optimization/Inkling-0.6B-A0.6B", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("inference-optimization/Inkling-0.6B-A0.6B")

input_ids = tokenizer("According to all known laws", return_tensors="pt").input_ids.to(model.device)
output = model.generate(input_ids, max_new_tokens=20)
print(tokenizer.decode(output[0]))

Creation Process

This model was created using the llm-compressor create-tiny-model claude skill.

Config inspected via inspect_config.py
Tiny model created via modified save_tiny_model.py — all-zero params fixed post init_weights
Fine-tuned on copypasta dataset; reached perplexity 1.45 (target: ≤3.0) at lr=5e-4
Checkpoint structure validated against original HuggingFace index
Inference validated via validate_tiny_model.py

Notes

The embed_tokens weights require explicit re-initialization after init_weights() (they initialize to zero in this architecture). The save script applies a fixup: any all-zero, non-finite, or extreme-valued parameter is re-initialized with kaiming_uniform / normal / ones as appropriate.
MTP (Multi-Token Prediction) layers present in the original checkpoint (model.mtp.*) are not included, as InklingForConditionalGeneration does not expose them through its standard interface.
Validation output: Success: 1.4451 <= 10.0