Thinking Machines Releases 975B Open Weights Inkling Model for Audio Transcription
Reactions from ranked influencers
2 posts
Thinking Machines' newly released 975B parameter open weights model, Inkling, supports transcription and ranks second among open weights models on AA-WER at 3.5% (#10), only behind Mistral's 24B Voxtral Small In addition to text and image input, Inkling accepts audio input, supporting use cases including speech transcription. We tested the 256K context window variant of the model, served via Thinking Machines' Tinker platform API. Model weights are available on Hugging Face under an Apache 2.0 license. Key takeaways ➤ Accuracy: Inkling scores 3.5% on AA-WER, ahead of transcription-focused open weights models like Voxtral Mini Transcribe 2 (3.6% WER, with 4B parameters), behind only Voxtral Small (2.8%, 24B). At 975B total parameters (41B active), Inkling is by far the largest open weights model to support transcription on our leaderboard ➤ Speed: Inkling processes audio at ~11x real-time - slower than dedicated transcription models such as Voxtral Mini Transcribe 2 (~81x) and MAI-Transcribe-1.5 (~264x) ➤ Cost: Inkling is available at $6.60 per 1,000 minutes of audio via Thinking Machines' Tinker platform, above transcription-focused models such as Voxtral Small ($4.00) and Voxtral Mini Transcribe 2 ($3.00) See more details below ⬇️
Inkling achieves a WER of 3.1% on AA-AgentTalk, 2.3% on VoxPopuli-Cleaned-AA, and 5.4% on Earnings22-Cleaned-AA.