Thinking Machines just released Inkling, its first major language model under an enterprise-friendly Apache 2.0 open source license, and it’s betting that “open” is not enough. The company built Inkling to be natively multimodal, meaning it can reason across text, images, and audio, and it also markets a specific behavior: it was designed “to answer directly on topics that may be subject to censorship.” For enterprises that want to run agentic AI workloads on-premises or in virtual private clouds, that combination matters because it targets two board-level anxieties at once: operational control and output constraints.

Inkling arrives with 975 billion total parameters and 41 billion active parameters, available on Hugging Face and through Thinking Machines’ own model training API, Tinker. On third-party benchmarks, it posts high, if sub state-of-the-art, performance for open weights models, including 77.6% on SWE-bench Verified, where it beats Nvidia Nemotron 3 at 71.9%, and 91.4% on VoiceBench, compared to 94.4% for Gemini 3.1 Pro on high reasoning effort. It also has a lighter sibling, Inkling-Small, previewed as a 276-billion-parameter alternative optimized for low latency and cost.

Now for the hook inside the hook: Inkling is built around a mechanism Thinking Machines calls “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. The company frames this as shifting along a cost/performance curve, including the claim that teams can reach the same score with a fraction of the tokens. In practical enterprise terms, this is a lever you can pull per workflow, instead of treating every request like it deserves full compute treatment. Lower the thinking effort for simpler tasks to reduce token spend, or crank it up for complex, multi-step reasoning where latency and cost are already part of the risk calculus.

This matters because the open-weight landscape is getting crowded, and not just with bigger numbers. Thinking Machines positions Inkling as a broad, balanced generalist rather than a single “god model” optimized for dominating every leaderboard. The benchmarks show what that means. On Design Arena’s Agentic Web Dev leaderboard (human frontend design scores), Inkling lands near the middle high-end at 1257, even as specialized labs produce models with elite reasoning and coding capabilities that outperform generalists on those dimensions. The source also notes a “fiercely competitive 2026 open-weight landscape” of highly specialized MoE architectures, where mixture-of-experts design can shift performance tradeoffs in interesting ways.

In that context, Inkling’s comparison set is doing a lot of work. GLM 5.2, described as widely considered the top open-weight reasoning model in the benchmark set, outperforms Inkling on pure coding, agentic, and complex reasoning tasks, including 62.1% on SWEBench Pro (Public) versus Inkling’s 54.3%, and 82.7 on Terminal Bench 2.1 versus Inkling’s 63.8. DeepSeek V4 Pro beats Inkling on SWEBench Verified (80.6% vs 77.6%) and SimpleQA Verified (57.0% vs 43.9%), but Inkling overtakes it on AIME 2026, hitting 97.1% versus DeepSeek’s 96.7%. Kimi K2.6 posts higher scores than Inkling on GPQA Diamond (91.1% vs 87.9%), BrowseComp (83.2% vs 77.1%), and HLE with tools (54.0% vs 46.0%), while Inkling scores higher on general chat instruction following, 79.8% on IFBench versus 76.0%. Versus the primary U.S.-based open-weight competition, Inkling’s parity and frequent superiority show up against Nemotron 3 Ultra, with 97.1% on AIME 2026 and 77.6% on SWE-bench Verified, beating Nemotron’s 94.2% and 70.7%, plus a big agentic jump on MCP Atlas, 74.1% versus 44.7%.

But the differentiator Thinking Machines is trying to sell hardest is not just performance. It’s epistemics and censorship behavior, including calibration, instruction following, and resistance to censorship. In an ecosystem where open weights can swing between overly restrictive safety guardrails and ideological boilerplate, Thinking Machines says Inkling was intentionally trained to answer directly on politically sensitive or heavily censored topics. The source says Thinking Machines validated this approach by submitting Inkling to the Propaganda and Censorship Eval developed by Cognition, where the published findings reported “strong patterns of censorship non-compliance,” described as resisting ideological capture or boilerplate refusals when presented with sensitive subjects. The model still includes defenses against genuinely malicious, dangerous, or illegal queries, and it was also evaluated on the StrongREJECT benchmark, which tests responses to una... (the source cuts off there), but the framing is clear: censorship resistance is not pitched as “anything goes,” it is pitched as “answer, unless the request crosses into truly harmful territory.”

Second-order implication: open-weight adoption is often sold internally as “we can control it,” but control is more than hosting. It becomes about what the model will refuse, when it will refuse, and whether those refusal patterns align with your organization’s risk policy and operational needs. A model engineered to compress internal reasoning steps through “chain of thought condensation” during large-scale reinforcement learning over 30 million rollouts, while also offering a tunable thinking budget, gives enterprises knobs for cost and latency that are typically missing from more rigid deployments. Pair that with explicit censorship-resistance positioning under an Apache 2.0 license, and you get a candidate that could shift procurement conversations away from closed-model black boxes toward customizable deployments that still behave the way compliance teams expect.

At the same time, the competitive reality is that closed models maintain a commanding lead in some domains. The source notes Claude Fable 5 hitting 95.0% on SWEBench Verified and 53.3% on HLE (text only), far above Inkling’s 77.6% and 30.0%, and GPT 5.6 Sol scoring 89.5 on Terminal Bench 2.1 versus Inkling’s 63.8. In native multimodality, Inkling competes but does not lead: 73.3% on MMMU Pro (Standard 10) versus Claude Fable 5’s 84.2% and GPT 5.6 Sol’s 83.0%, and 77.2% on MMAU versus Gemini 3.1 Pro’s 82.5%. So the strategic stakes for executives are not that Inkling beats everyone at everything. It’s that Thinking Machines is trying to carve out a distinct enterprise wedge: most capable open-weights foundation model that natively fuses text, vision, and audio, while giving developers direct programmatic control over the cost-to-performance ratio, plus a stance on answering censored topics. In a market where agentic AI procurement is moving fast, that wedge could be enough to earn pilots, then deployments.