Inkling – Open-Weights 975B Parameter LLM
TL;DR Highlight
Thinking Machines released Inkling, an open-weights MoE model taking text, image, and audio input.
Who Should Read
Backend developers comparing open-weights models to fine-tune on their own domain data
Core Mechanics
- Inkling activates 41B of 975B
- It is a Mixture of Experts model, where only some expert blocks run per input.
- Total parameters are 975B and active parameters are 41B.
- Inkling-Small is a quarter of the size
- It has 276B total and 12B active parameters.
- The post says it matches the larger sibling on many benchmarks with lower cost and latency.
- Text, image, and audio input
- Both models accept text, images, and audio.
- The post says it is at the open-source frontier for speech.
- 1M context, shorter on Tinker
- The model spec lists 1M tokens of context.
- Tinker offers 64K or 256K.
- Thinking effort is adjustable
- Controllable effort trades speed against performance.
Community Reactions
- One commenter shared the model card link and the 975B total, 41B active summary
- Detailed specs are on the separate model-card page.
- A moderator noted this is a duplicate submission
- Another HN thread for the same post already exists.
How to Apply
- Read the Hugging Face model card first
- Today, check the license, benchmark scores, and supported input types.
- Compare its context length with your needs.
- Shortlist Inkling-Small first
- If latency or cost matters, run your own eval set on Small.
- Measure whether the gap to the larger model is acceptable.
- Trial native multimodal input
- If you convert audio or images to text first, feed the raw input instead and compare results.
- Account for Tinker context limits
- If you plan to use Tinker, design input chunking around 64K/256K context.
Terminology
Related Papers
Show HN: Qwen3-4B with the AI writing tics trained out
Distilling Qwen3-4B with LoRA cut its AI-writing slop score from 121 to 12.5 per 1000 words
Dust: Pretraining Transformers Without Backpropagation
Adding per-token noise to activations lets transformers pretrain without backpropagation, rivaling backprop.
Show HN: TurboGPT: train 22KiB transformer in 13s
A CUDA C++ project trains a tiny byte-level GPT in under a minute.
I had Gemini train its own replacement for $9
Fine-tuning GLiNER on $9 of Gemini labels reached 0.83 F1 locally.
Learning to solve hard problems in RL for LLMs by never giving up
Cutting compute on easy problems and retrying hard ones in RL raised accuracy on the hardest problems
Training a 3.8B LLM to 0.384 CORE for $998
A 3.8B LLM trained on 8 B200s in 43 hours for $998 scored 0.384 on CORE.
I were 17, I'd learn how to build LLMs from scratch