Thinking Machines open-sources Inkling-2 Mini: 26.4B MoE with native vision and 1M context
solyarisoftware · x · 2026-11-22
Thinking Machines released Inkling-2 Mini under Apache 2.0, and early users are calling it the model many were waiting for from Qwen4-35B:
- Scale: 26.4B total parameters with only 3.4B active per token (5.4B effective activations), a 128-expert MoE picking 8 per token, with the middle 12 layers run 3 times
- Capabilities: native vision (text + images in, text out) and up to 1M token context
- Local deployment: MoE experts can be offloaded to RAM, so it runs on 8GB VRAM + 16GB RAM machines; GGUF builds include 11.7GB IQ3XXS and 17.6GB Q4KXL
The poster runs the 17.6GB Q4KXL on an RTX 3090 and preferred its one-shot outputs over Hy4-flash.
Related event: Thinking Machines releases Inkling 2 Mini, a 26B recurrent transformer(116 posts)→
More from Models
- Defending GPT-6.2: a detail-obsessed 'trees' model built for agents, not prose — koltregaskes · 2026-11-22
- Tracking how Claude's utopian visions have evolved over time — amplifiedamp · 2026-11-22
- LLMs have always self-corrected: Opus 5.5 caught its own mistake mid-generation in real time — amplifiedamp · 2026-11-22
- Open-weight models can no longer be suppressed: Pirate Face turns model weights into immortal torrents — esrtweet · 2026-11-22
- 'Compression of the RH proof has begun': AI circles buzz over unverified Riemann claim — teortaxesTex · 2026-11-22
- GLM 5.3 and DeepSeek V4.1 Flash resold at 60% off: the economics of open-source giveaways — lxfater · 2026-11-22