Uploaded July 2026 | Updated November 2026, 4 days ago
In this video, I break down Inkling, Thinking Machines’ first open-weight model release under Apache 2.0, trained from scratch as a multimodal decoder-only MoE. I cover the core architecture (near‑trillion parameters, 1M context window, 45T multimodal tokens, 256 experts with 41B active params), how it compares on benchmarks (not SOTA overall but a strong generalist with notable design performance), and why it matters as a major Western open model release. I also walk through the Tinker platform for fine-tuning/post-training, mention the upcoming Inkling Small (276B/12B active), show playground settings like reasoning effort, context limits, and web search/tool calling, demonstrate website generation and a coding example (ISS tracker), and review pricing and options to run weights yourself.
LINK:
thinkingmachines.ai/news/introducing-inkling
My voice to text App: whryte.com
Website: engineerprompt.ai
RAG Beyond Basics Course:
prompt-s-site.thinkific.com/courses/rag
Signup for Newsletter, localgpt:
https://tally.so/r/3y9bb0
Let's Connect:
🦾 Discord: discord.com/invite/t4eYQRUcXB
☕ Buy me a Coffee: ko-fi.com/promptengineering
|🔴 Patreon: patreon.com/PromptEngineering
💼Consulting: calendly.com/engineerprompt/consulting-call
📧 Business Contact: engineerprompt@gmail.com
Become Member: tinyurl.com/y5h28s6h
💻 Pre-configured localGPT VM: bit.ly/localGPT (use Code: PromptEngineering for 50% off).
Signup for Newsletter, localgpt:
https://tally.so/r/3y9bb0
Inkling: Thinking Machines’ First Open-Weight Multimodal MoE Model (Architecture, Playground Demos & Pricing)
00:00 Inkling
00:26 Why This Matters
01:05 Architecture and Scale
01:55 Inkling Small
02:57 Multimodal Strengths
04:06 Reasoning Effort Controls
04:36 Playground Tour and Prompt
06:50 Tool Use and Speed
08:13 Website Design Patterns
09:21 Coding Demo ISS Tracker
In this video, I break down Inkling, Thinking Machines’ first open-weight model release under Apache 2.0, trained from scratch as a multimodal decoder-only MoE. I cover the core architecture (near‑trillion parameters, 1M context window, 45T multimodal tokens, 256 experts with 41B active params), how it compares on benchmarks (not SOTA overall but a strong generalist with notable design performance), and why it matters as a major Western open model release. I also walk through the Tinker platform for fine-tuning/post-training, mention the upcoming Inkling Small (276B/12B active), show playground settings like reasoning effort, context limits, and web search/tool calling, demonstrate website generation and a coding example (ISS tracker), and review pricing and options to run weights yourself.
LINK:
thinkingmachines.ai/news/introducing-inkling
My voice to text App: whryte.com
Website: engineerprompt.ai
RAG Beyond Basics Course:
prompt-s-site.thinkific.com/courses/rag
Signup for Newsletter, localgpt:
https://tally.so/r/3y9bb0
Let's Connect:
🦾 Discord: discord.com/invite/t4eYQRUcXB
☕ Buy me a Coffee: ko-fi.com/promptengineering
|🔴 Patreon: patreon.com/PromptEngineering
💼Consulting: calendly.com/engineerprompt/consulting-call
📧 Business Contact: engineerprompt@gmail.com
Become Member: tinyurl.com/y5h28s6h
💻 Pre-configured localGPT VM: bit.ly/localGPT (use Code: PromptEngineering for 50% off).
Signup for Newsletter, localgpt:
https://tally.so/r/3y9bb0
Inkling: Thinking Machines’ First Open-Weight Multimodal MoE Model (Architecture, Playground Demos & Pricing)
00:00 Inkling
00:26 Why This Matters
01:05 Architecture and Scale
01:55 Inkling Small
02:57 Multimodal Strengths
04:06 Reasoning Effort Controls
04:36 Playground Tour and Prompt
06:50 Tool Use and Speed
08:13 Website Design Patterns
09:21 Coding Demo ISS Tracker




![CLM-8B: Contrastive Language Models
Jev made System 1 models hot. CLM from Stanford and NVIDIA Research takes a completely different approach: it never generates an answer.
Instead, it embeds the situation and every possible action into the same space and picks whichever action lands closest. Because actions are embedded once and cached, 1,000 options cost about the same as one: 44 ms vs 579 ms for Jev.
The catch: each option is encoded on its own, so it never compares them side by side.
Full breakdown + demos on my DGX Spark: [LINK TO FULL VIDEO]
📄 Blog: https://contrastive-lm.notion.site
💻 Code: https://github.com/Contrastive-LM/CLM
#AI #LLM #MachineLearning #SystemOne #Jev #CLM #NVIDIA CLM-8B: Contrastive Language Models](https://i.ytimg.com/vi/MK1zLrR9gpM/mqdefault.jpg)





