Welcome Inkling by Thinking Machines
Thinking Machines releases Inkling, an open-weight ~1T-parameter (41B active) Mixture-of-Experts multimodal model with a 1M context window, trained on 45T tokens of text, images, audio and video, with day-one support across Transformers, vLLM, SGLang, llama.cpp and Unsloth.
Marked critical by the editors; it stays in this band.
On this page12 sections
The editors' notes below — why it matters, the context, the learning path — are written in Chinese.
Record
The signal
#Inkling is a large open-source multimodal language model with roughly 1 trillion total parameters and 41 billion active parameters via a 256-expert Mixture-of-Experts architecture. It natively processes image, text and audio inputs through simple hierarchical MLP modules rather than separate encoders, supports agentic coding and reasoning, ships speculative multi-token prediction (MTP) layers for faster inference, combines global and sliding-window attention, and offers adjustable reasoning effort from none to max. Weights are released on Hugging Face in BF16 (~2TB VRAM) and a quantized NVFP4 variant (~600GB VRAM).
Why now
Why it matters
#开源权重第一次同时具备万亿参数、原生多模态与 1M 上下文这组旗舰规格，直接改变「开源 vs 闭源 API」的选型格局；而 600GB 起步的显存需求也把自托管的容量规划难度抬到了新档位。
Background
Technical context
#MoE 稀疏激活让 1T 总参数只需 41B 激活算力，是「大容量、可控推理成本」的主流路线。但显存必须装下全部专家权重：即便 NVFP4 量化后仍需约 600GB，对应 8×80GB 级别的多卡甚至多机部署——权重开放不等于部署门槛消失。
Audience
Who should care
#Next
Learning path
#- 先理解 MoE 的稀疏激活机制：总参数、激活参数、显存占用三者为什么脱钩
- 对照 BF16 / NVFP4 两档显存需求，估算自托管 Inkling 的硬件与成本下限
- 用 Hugging Face Inference Providers 的免费额度在自己的任务上先做质量评估，再决定是否投入部署
Network
Where it sits in the network
#How this signal connects to neighbouring technologies, skills and background knowledge. Follow any link to keep exploring.
Further thought
Follow-up questions
#- NVFP4 量化相对 BF16 的质量损失在你的任务上有多大？
- vLLM / SGLang 对 MTP 投机解码层的支持成熟度如何？
- 1M 上下文在真实长文档任务中的有效利用率是多少？
Publication record
Featured in
#The briefings that carried this signal, newest first; the last row is the weekly review of its week. The tag is the section it ran in.
技术简报 - 2026-07-19
Immediate attentionWeekly review · Jul 13 – Jul 19
Immediate attention
Sources