Thinking Machines, the American AI startup founded by former OpenAI CTO Mira Murati, has released Inkling, its first major multimodal language model under an enterprise-friendly Apache 2.0 open-source licence. This release marks a significant milestone for enterprises seeking to deploy customisable, on-premises AI workloads without the constraints of proprietary systems. The model boasts impressive performance on software engineering benchmarks, achieving 77.6% on SWE-bench Verified, surpassing competitor Nvidia Nemotron 3's 71.9%. Additionally, Inkling demonstrates strong voice understanding capabilities with 91.4% on VoiceBench.
What sets Inkling apart is its 975 billion total parameters operating as a Mixture-of-Experts system, where only 41 billion parameters activate during token generation. The model natively processes text, images, and audio through an innovative encoder-free early fusion approach, eliminating the need for bolted-on external encoders. Perhaps most notably, Thinking Machines designed Inkling with "controllable thinking effort," allowing developers to programmatically adjust the model's reasoning budget from 0.2 to 0.99. This feature enables organisations to balance cost against performance, using fewer tokens for simpler tasks whilst scaling up computational power for complex, multi-step reasoning challenges.
A particularly bold aspect of Inkling's design is its explicit resistance to censorship. The model was intentionally trained to answer directly on politically sensitive or heavily censored topics, validated through Cognition's Propaganda and Censorship Eval. Despite this openness, Inkling maintains robust safety standards, scoring 98.6% on the StrongREJECT benchmark for refusing genuinely harmful requests. On the FORTRESS benchmark, it achieved a 78.0% refusal rate on adversarial queries involving weapons or violence, whilst maintaining 95.9% compliance on legitimate requests. Thinking Machines recommends deploying additional moderation tools like Llama Guard for application-level safety.
Whilst Inkling doesn't claim state-of-the-art status across all benchmarks, it occupies a unique position in the competitive 2026 landscape. Chinese models like GLM 5.2 and DeepSeek V4 Pro outperform it in specific coding and reasoning domains, whilst closed-source systems like Claude Fable 5 and GPT 5.6 Sol maintain leads in peak reasoning capabilities. However, Inkling distinguishes itself as the most capable open-weight foundation model that natively fuses text, vision, and audio whilst offering programmatic cost-performance control. The Apache 2.0 licence provides true open-source freedom, allowing developers to download, modify, and commercialise the model entirely royalty-free—a stark contrast to many "open" models tethered to commercial restrictions or revenue caps.
Artículo generado mediante LaRebelionBOT
No hay comentarios:
Publicar un comentario