Enterprises looking to move more of their agentic AI workloads to open weights models they can customize, control and run on-premises or in virtual private clouds have a strong new contender to consider.
Today, Thinking Machines—the highly capitalized American AI startup founded by former OpenAI CTO Mira Murati—released Inkling, its first major language model under an enterprise-friendly Apache 2.0 open source license, and it boasts high, if sub state-of-the-art, performance for open weights models on third-party benchmarks, specifically software engineering (77.6% on SWE-bench Verified, where it beats fellow U.S. open rival Nvidia Nemotron 3’s 71.9%) and voice understanding (91.4% on VoiceBench compared to 94.4% for Gemini 3.1 Pro on high reasoning effort).
Another differentiator: Thinking Machines notes that Inkling was designed “to answer directly on topics that may be subject to censorship,” offering enterprises concerned about factual outputs, irrespective of controversy or sensitivity, a more trustworthy option.
Coming in at 975 billion total parameters, Inkling is a natively multimodal, open-weights Mixture-of-Experts (MoE) system capable of reasoning across text, images, and audio. The weights are already available on Hugging Face and the company’s own model training application programming interface (API), Tinker.
Designed to balance cost against performance through a novel “controllable thinking effort” mechanism, the model represents a significant departure from the black-box scaling strategies of frontier competitors.
Alongside the flagship model, Thinking Machines also announced a preview of Inkling-Small, a lighter 276-billion-parameter alternative optimized for workloads where low latency and cost are paramount.
While Inkling is a formidable multimodal engine, it lands in a fiercely competitive 2026 open-weight landscape characterized by highly specialized MoE architectures. Rather than attempting to dominate every leaderboard, Thinking Machines explicitly designed Inkling—with 975 billion total and 41 billion active parameters—as a broad, balanced generalist.
For example, it comes in near the middle high-end of benchmark performance 1257 on Design Arena’s Agentic Web Dev leaderboard measuring human scores of frontend web design.
But China’s leading AI labs have produced models with elite reasoning and coding capabilities, posing a stiff challenge to Inkling’s generalist approach and ultimately outperforming it on general and coding benchmarks.
GLM 5.2: Widely considered the top open-weight reasoning model available in the benchmark set, GLM 5.2 outperforms Inkling on pure coding, agentic, and complex reasoning tasks. It scores 62.1% on SWEBench Pro (Public) compared to Inkling’s 54.3%, and a massive 82.7 on Terminal Bench 2.1 against Inkling’s 63.8. GLM 5.2 also holds the edge in text-only reasoning, scoring 40.1% on HLE (text only) versus Inkling’s 30.0%.
DeepSeek V4 Pro: DeepSeek maintains an edge in several strict coding and factuality domains, beating Inkling on SWEBench Verified (80.6% vs. 77.6%) and SimpleQA Verified (57.0% vs. 43.9%). However, Inkling successfully overtakes DeepSeek V4 Pro in mathematical problem-solving, achieving 97.1% on AIME 2026 compared to DeepSeek’s 96.7%.
Kimi K2.6: This model outpaces Inkling across multiple technical benchmarks, delivering higher scores on GPQA Diamond (91.1% vs. 87.9%), BrowseComp (83.2% vs. 77.1%), and HLE with tools (54.0% vs. 46.0%). Yet Inkling proves more resilient on general chat instruction following, scoring 79.8% on IFBench compared to Kimi K2.6’s 76.0%.
Against its primary U.S.-based open-weight competition, Inkling demonstrates strong parity and frequent superiority.
Nemotron 3 Ultra: Inkling consistently outperforms this U.S. rival across reasoning and coding. Inkling posts 97.1% on AIME 2026 and 77.6% on SWEBench Verified, beating Nemotron’s 94.2% and 70.7%, respectively. Furthermore, Inkling significantly leads in agentic workflows, scoring 74.1% on MCP Atlas against Nemotron’s 44.7%.
When compared to closed-source juggernauts like Claude Fable 5, GPT 5.6 Sol, and Gemini 3.1 Pro, Inkling trails in peak reasoning and software engineering autonomy, but remains highly competitive in multimodality.
Coding and Reasoning: Closed models maintain a commanding lead. Claude Fable 5 (max) hits 95.0% on SWEBench Verified and 53.3% on HLE (text only), far outpacing Inkling’s 77.6% and 30.0%. GPT 5.6 Sol dominates Terminal Bench 2.1 with an 89.5, easily clearing Inkling’s 63.8.
Native Multimodality: Inkling’s native visual and audio capabilities hold their own. On the MMMU Pro (Standard 10) vision benchmark, Inkling’s 73.3% is competitive, though trailing Claude Fable 5’s 84.2% and GPT 5.6 Sol’s 83.0%. In audio processing, Inkling scores a highly respectable 77.2% on MMAU, keeping it within striking distance of Gemini 3.1 Pro’s 82.5%.
If an enterprise workflow demands elite software engineering autonomy or the highest bounds of text-only reasoning, models like GLM 5.2 or proprietary systems like Claude Fable 5 maintain the edge.
However, Inkling carves out a unique and highly defensible position: it is the most capable open-weight foundation model that natively fuses text, vision, and audio, while simultaneously offering developers direct programmatic control over the cost-to-performance ratio.
Rather than attempting to build a singular “god model” optimized strictly for state-of-the-art benchmark domination, Thinking Machines engineered Inkling for adaptability and efficiency in real-world workflows.
The standout feature of this release is Inkling’s “controllable thinking effort.” Developers can programmatically adjust the model’s reasoning budget—scaling from 0.2 to 0.99—to dictate how hard the AI should “think” before generating an output.
As the company noted, “Inkling’s continuous thinking effort lets you pick your point on the cost/performance curve—reaching the same score with a fraction of the tokens”.
In practical terms, this allows enterprises to deploy Inkling with lower token expenditure for simpler tasks, while cranking up the compute overhead for complex, multi-step reasoning challenges. However, by keeping the thinking effort lower and generating fewer tokens, the cost-conscious enterprise can achieve high quality results and performance on simple tasks while spending less money, or, in the case of those running models locally, less costs on energy and compute resources.
During the model’s large-scale reinforcement learning (RL) training over 30 million rollouts, researchers observed an emergent phenomenon they called “chain of thought condensation”. Over time, Inkling naturally learned to compress its internal reasoning steps—dropping grammatical overhead and connectives—while reaching the same accurate conclusions, resulting in drastically reduced latency.
A notable element of Thinking Machines’ release is its explicit focus on the model’s epistemics—specifically its calibration, instruction following, and resistance to censorship.
In an ecosystem where open-weight models adopt either overly restrictive safety guardrails or echo state-aligned ideological talking points, Inkling was intentionally trained to answer directly on politically sensitive or heavily censored topics.
To validate this approach, Thinking Machines submitted Inkling to the Propaganda and Censorship Eval developed by AI startup Cognition. According to the published findings, Inkling demonstrated “strong patterns of censorship non-compliance,” effectively resisting ideological capture or boilerplate refusals when presented with sensitive subjects.
Despite its resistance to censorship, the model maintains a robust defense against genuinely malicious, dangerous, or illegal queries. On the StrongREJECT benchmark—which tests responses to unambiguous harmful requests—Inkling scored 98.6%, placing it in line with strict frontier safety standards. Furthermore, on the FORTRESS benchmark, Inkling successfully navigated the line between safety and over-refusal: it achieved a 78.0% refusal rate on adversarial queries (such as those involving weapons, cyberattacks, or violence) while maintaining a 95.9% compliance rate on benign, look-alike queries.
Thinking Machines noted that typical open-weight vulnerabilities remain within the architecture. Internal safety evaluations revealed an “occasional tendency to comply with role-play and indirectly framed prompts concerning harmful topics”. The company advised enterprise developers to treat the model’s built-in refusals as just one layer of security, recommending the downstream deployment of external moderation tools—such as Llama Guard—to filter adversarial jailbreaks and enforce use-case-specific safety policies at the application level.
Inkling’s scale is staggering, yet sparse. The MoE architecture features 975 billion total parameters, but only 41 billion parameters are active during any given token generation. It supports a massive context window of 1 million tokens and diverges from typical transformer models by using relative positional embeddings instead of the industry-standard Rotary Positional Embedding (RoPE).
True to the company’s foundational vision, Inkling was trained from scratch to be natively multimodal. Unlike models that rely on bolted-on external encoders, Inkling uses an encoder-free early fusion approach. It directly ingests audio as discrete dMel spectrograms and visual data as 40×40 pixel patches via a hierarchical multi-layer perceptron (hMLP), projecting all modalities into a shared hidden space.
For enterprise IT teams and developers, the most disruptive aspect of Inkling may be its licensing. Inkling is released under the permissive Apache 2.0 license.
In an ecosystem where many so-called “open” models from Western labs are tethered to dual-use commercial licenses, acceptable use restrictions, or revenue caps, an Apache 2.0 designation makes Inkling a true open-source foundation. This gives developers the legal freedom to download, modify, integrate, and commercialize the model weights entirely royalty-free.
The model is readily deployable across major open-source inference libraries—including SGLang, vLLM, TokenSpeed, and llama.cpp—and comes with a native NVFP4 quantized checkpoint optimized for NVIDIA Blackwell systems.
The AI community’s response has been swift, praising both the model’s openness and the underlying engineering execution.
In a post on X, Thinking Machines co-founder John Schulman reflected on the rapid development cycle: “Inkling is out today, with open weights and in Tinker. It’s been fun to watch this one come together: pretraining began last winter, and starting in mid-January a small team built up the coding, reasoning, and agentic training from there. We learned a lot building it, and I hope people find good uses for it.”
Horace He, a researcher at Thinking Machines (previously from PyTorch), underscored the difficulty of the task in another post on X: “It truly takes a village to release a model, perhaps especially an open weights model. Actually doing the entire process from scratch, from data to pretraining to posttraining to actual release, gives a lot of appreciation for anyone who does it!”
The broader open-source ecosystem has also embraced the technical integrations. Lysandre Debut, the Chief Open-Source Officer at Hugging Face, shared his enthusiasm regarding the model’s optimization in his own X post: “One thing I find quite striking is how much easier accelerating models has become… We replaced the model’s causal Conv1D with the `causal-conv1d` kernel. One line changed, +4% tokens per second. We then replaced its attention implementation with FlashAttention-4. Another single change, another +11%. That’s a total throughput improvement of about 15%, without changing the model architecture or retraining anything.”
Tiezhen Wang, an ecosystem growth expert and ex-Googler, celebrated the release as a massive win for the open-source community, listing the model’s impressive specifications on X, highlighting its “975B total, 41B active” size, “Native MTP support,” and the highly coveted “Apache 2.0 license.”
To understand the significance of Inkling, one has to look back at the rapid trajectory of Thinking Machines over the past 18 months.
When Mira Murati departed OpenAI in late 2024 to found Thinking Machines alongside industry veterans like John Schulman and Barret Zoph, the stated goal was to pivot away from building isolated autonomous agents. Instead, the company aimed to build flexible, multimodal systems designed for genuine human-AI collaboration and open science.
By July 2025, the startup had secured a historic $2 billion seed round led by Andreessen Horowitz at a $12 billion valuation. At the time, Murati promised the impending release of a product with a “significant open source component” to empower researchers and startups.
The company’s philosophy began coming into sharper focus in October 2025 with the launch of Tinker, a Python-based API for large language model fine-tuning that gave researchers granular control over training pipelines without the friction of distributed compute management.
That same month, Thinking Machines researcher Rafael Rafailov delivered a provocative critique of the AI industry at TED AI. He argued that the current trajectory of simply throwing more compute at models was fundamentally flawed, noting that today’s systems take shortcuts—like wrapping code in try/except blocks—because they are trained strictly for task completion rather than genuine learning.
Rafailov posited that the first artificial superintelligence would not be a “god model,” but rather a “superhuman learner” capable of meta-learning and internalizing abstractions. Inkling’s architecture—specifically its controllable thinking effort and its ability to organically compress its chain of thought during RL—feels like the first tangible realization of Rafailov’s thesis.
In May 2026, the lab teased its technical prowess with the research preview of TML-Interaction-Small, a system that eliminated “turn-based” chat by processing inputs and outputs simultaneously in 200ms chunks. This “full-duplex” breakthrough proved the company could build highly responsive, natively multimodal models from scratch.
Now, with Inkling out in the wild, Thinking Machines has delivered on its foundational promises. By offering a massive, natively multimodal model under a true open-source license, they aren’t just giving developers a new tool—they are attempting to fundamentally rewrite the economics and accessibility of frontier AI development.
The most iconic esports player ever has signed a multi-year deal to sit alongside Cristiano Ronaldo and Magnus Carlsen as EWC ambassadors.
For the first time in history, we have access to the right technology to tackle the insider risk problem.
A new Pixel 11 price leak flips Google’s usual advantage over Apple. Here’s what Google is betting on to convince you anyway.
The enterprise AI industry has a math problem. Cisco data shows 85% of enterprises are piloting AI agents, but only 5% have shipped them to production. At VB Transform 2026 on Tuesday, Bryan Silverthorn, Director of AGI Autonomy at Amazon, explained why that gap persists — and why the answer isn’t better benchmarks.
Silverthorn, who joined Amazon through its acquisition of Adept AI and now leads multimodal agent training inside the company’s AGI lab, argued that reliability must be broken into four distinct dimensions: consistency, robustness, predictability, and safety — a framework he credits to research from Princeton.
“It unpacks different factors that I see tangled together in almost every eval I’ve ever seen,” he said.
The framework matters because agents routinely ace internal evaluations and then collapse in the wild. Silverthorn described a customer that deployed an agent for software QA involving serial number extraction from screens. It worked flawlessly for two months — then began intermittently reading wrong numbers. The culprit: the underlying vision encoder behaved differently depending on where the serial number appeared on screen, and a software change imperceptible to humans triggered the failure.
The lesson, Silverthorn said, is about measurement, not just models. “The models have to be better. Obviously, we’re working hard on making the models better,” he said. But the deeper takeaway, he added, is that teams need to identify their dimensions of variability and match measurement rigor to the stakes of the application. VentureBeat’s own proprietary research, presented before the session, reinforces the point: half of surveyed companies shipped agents that passed internal evals but failed real customers, and enterprises overwhelmingly track uptime while ignoring accuracy — checking the pulse without checking the diagnosis. A related finding underscored how few guardrails exist: most enterprises default to the model makers’ own evaluations and little else, leaving their testing strategy, as I described it on stage, a coin flip between trusting the vendor and trusting nothing.
Silverthorn’s most memorable prescription was cultural, not technical. Inside Amazon’s AGI lab, researchers literally call their agents “interns” — as in, “I’ll have my intern talk to your intern.” The joke carries a serious operational philosophy. Agents, like interns, are powerful but occasionally clueless, capable of amazing work and spectacular derailment.
Managing them, he argued, requires management skills rather than software skills: asking what could go wrong, adding backups and undo capabilities, and consciously deciding what risk you can accept. “You can ask the intern, ‘Hey, what might you do wrong here? How might you mitigate your negative outcomes?'” he said. Amazon’s lab has embraced that trade-off, accepting agents occasionally running the wrong experiment in exchange for research velocity — including one agent running experiments around the clock on its own high-level research plan.
Silverthorn was candid about the limits of today’s technology. Self-improving AI remains “a loaded term,” he said — Amazon uses AI to improve its models constantly, but fully autonomous self-improvement is distant. Computer use remains a core focus of his lab, with a commercial trucking customer already using browser automation to stitch together warranty claims across fragmented systems**, though he stressed that no future agent will rely on computer use alone — it will work alongside MCP, APIs, and other tools to complete end-to-end workflows**. And LLM-as-judge techniques, while promising, are just one of several strategies for aligning agent capability with acceptable risk.
For enterprises stuck in pilot purgatory, the path forward starts with a mindset shift: stop asking whether your agent can do something impressive once, and start asking whether it can do it correctly a thousand times in a row.
In other words, the enterprises that escape the 85% ceiling won’t be the ones with the smartest agents. They’ll be the ones with the best managers.
Hundreds of enterprise leaders and technical experts packed the main ballroom of the luxurious Hotel Nia in Menlo Park this week for VB Transform 2026, the year’s preeminent conference on using generative AI agents to drive business outcomes.
Rachad Alao, vice president of product engineering at the rising Canadian enterprise AI startup Cohere, joined VentureBeat CEO and editor-in-chief Matt Marshall for a fireside chat about building agentic systems without surrendering sensitive data, infrastructure control, or the ability to change vendors.
Alao, who previously led responsible AI and trust and safety engineering teams at Google and Meta, argued that AI sovereignty means more than downloading an open model or running an application behind a corporate firewall.
Asked how Cohere defines sovereignty, Alao pointed to organizations operating mission-critical systems, including banks, hospitals and governments.
“It is important to have very tight control on where the data resides, have tight control on the AI,” he said, adding that AI operations should take place in jurisdictions an organization understands or directly controls.
That extends from GPUs and private-cloud infrastructure through governance systems that route requests among models, as well as the connectors, search tools and agent frameworks acting on enterprise data.
“You want to have control on the entire stack,” Alao said.
Marshall challenged one of the central economic arguments for smaller, locally deployed models: Inference prices continue to fall rapidly, potentially weakening the case for optimizing every token.
Alao countered that total consumption is climbing even faster as enterprises move from relatively simple chatbots to agents that reason through problems, call tools, search internal systems and take multiple steps before returning an answer.
“Your token utilization is going exponentially up, because you’re dealing with more and more complex agentic use cases,” he said. Those workflows require “a lot of processing, thinking, tools interaction” to complete their objectives, he added.
Alao also drew a contrast between providers that bill customers according to token consumption and Cohere’s approach.
“If your whole way of charging customers is for token utilization, you want to maximize token utilization,” he said. “We do not sell our models and our platform that way.”
Instead, Alao said Cohere tries to help enterprises solve their hardest problems privately and securely while reducing unnecessary model usage. His prescription was straightforward: “Use the right model for the task at hand.”
Rather than sending every request to the largest available frontier model, enterprises should route work according to the intelligence required and the sensitivity or regulatory burden attached to the task.
Alao cited an unnamed Canadian bank that uses Cohere’s on-premises models for highly regulated workloads, while sending less sensitive tasks requiring greater intelligence through Cohere’s North platform to larger frontier models.
“So model routing can become super useful,” he said.
Asked by an audience member how Cohere’s open-source North Mini Code, released last month, could compete against proprietary coding models, Alao acknowledged that larger frontier models may perform somewhat better on the hardest tasks.
But that advantage may not justify using them indiscriminately.
“For 80% of the use cases that they needed, this was a lot more effective, a lot cheaper,” Alao said of developers adopting the model.
Cohere’s North Mini Code runs on a single Nvidia H100 GPU and targets agentic software engineering, including terminal work, code review and tool use.
The company has also released Command A+, a 218-billion-parameter mixture-of-experts model with only 25 billion parameters active during each generation step.
Its compressed four-bit version reduces the hardware required for private deployment, while its Apache 2.0 license gives enterprises broad freedom to operate and modify it.
Asked about Cohere’s longstanding work on embeddings and enterprise search, Alao said the field is moving beyond retrieving text and inserting it into a model’s context window.
“Today, the state of the art is around multimodal search,” he said. “It’s beyond just the text modality.”
Search across documents, images and other forms of information is becoming “an integral component of your agentic workflow,” Alao added, with the model deciding when and how to use retrieval like any other tool.
Asked what would persuade enterprises to move beyond bundled AI services from existing cloud providers, Alao returned to data control and portability.
“If you’re interested in sovereignty, you want to have more control on your data,” he said. Cohere’s governance layer, he added, lets customers route traffic to appropriate models, “breaking that vendor lock-in concern that a lot of our customers have.”
Apple’s next-gen software introduces key upgrades to AirPods Pro 2, AirPods 4, AirPods Pro 3 and AirPods Max 2. The software’s now available for iOS 27 public beta users.
As operational complexity continues to increase, the challenge becomes less about access to data and more about making sense of it in the moments that matter.
Samsung has confirmed new Flex Titanium display tech for the Galaxy Z Fold 8, promising a smaller crease and better durability, but will repairs cost more or less?
While the technology has been available for years, the changing legal framework has played a major role in B2B adoption.