$40 Billion Valuation, US Company Finally Creates a Competitor to China's Open-Source Models
Author: Wildcard, GeekPark
On November 5, two major developments emerged for Thinking Machines Lab (TML), co-founded by OpenAI's former CTO Mira Murati.
That day, the company officially released its second-generation model, Inkling 2 Small. Third-party authoritative evaluation firm Artificial Analysis (AA) gave it high praise: after weights are fully released, it will become the "highest-scoring open-weight model outside China."
Meanwhile, market reports indicate that this star startup is raising a new round of $5 billion financing at a $40 billion pre-money valuation.
TML made its highly publicized debut in February 2025, marking over a year and a half since its founding. During this period, it secured a $2 billion seed funding round led by a16z, but also experienced turbulence and attrition within its core team—co-founder Andrew Tulloch moved to Meta, while co-founders Barret Zoph, Luke Metz, and others returned to OpenAI. Subsequently, PyTorch co-founder Soumith Chintala took over the role of CTO.
Reports in January indicated that fundraising negotiations previously underway at an approximate $50 billion valuation had stalled. Therefore, whether this round of high-valuation financing can successfully close largely depends on the hardcore capabilities of its second-generation model.
Reading through the technical blog for Inkling 2 Small, beyond the standard benchmark charts and architecture diagrams, the most thought-provoking section is a candid account detailing the data sources for training: data used in the latter half of the model's pretraining was largely sampled from earlier model trajectories, explicitly naming Moonshot AI's Kimi K3.
Where the 594B Small Excels
In today's generative AI industry, the parameter scale denoted by "Small" is constantly being redefined.
TML's first-generation open-weight model Inkling, released in July this year, features 975B total parameters, followed by Inkling Small at 276B two weeks later. In its second generation, Inkling 2 Small's total parameter count more than doubles, climbing directly to 594B, activating 17B parameters per inference. It utilizes a MoE (Mixture of Experts) architecture, natively supports 1 million token context and adjustable reasoning depth, supports visual multimodal inputs, but removes the direct audio input capability present in the first generation.
Its underlying training pipeline is extremely demanding. In the first stage of pretraining, leveraging NVIDIA's GB300 NVL72 rack clusters, the team processed 8.4 trillion tokens in 68 hours at native NVFP4 (4-bit floating point) precision. The second stage further injected 3.8 trillion tokens of high-quality incremental data. The final two stages concluded with approximately 354 billion tokens of complex reasoning and long agent trajectories, fully extending the context window to 1 million tokens.
This was immediately followed by nine days of large-scale joint reinforcement learning: utilizing 2,048 GPUs to complete over 17 million rollouts, with a median time of 8 minutes per rollout, and single long-horizon tasks running for up to 4 hours.
This rigorous, long-chain training aims precisely at what the official describes as the greatest breakthrough of this generation of models: reliable long-horizon autonomous working capabilities. In official technical demos, the model directly wrote a complete city management simulation game from scratch using Three.js. Its building meshes, terrain textures, and vehicle dispatch were entirely auto-generated by the program, while dynamically handling power supply, fiscal budgets, and random sudden disasters. Additionally, it could write a hardware simulator from scratch in Python capable of successfully booting a real Linux kernel, and even transform a 2D floor plan end-to-end into a 3D indoor roaming scene supporting free camera angles.
Figure: City management game simulation generated from scratch using Three.js
The entire RL training process was equally thrilling and precarious. Midway through RL, the backend monitoring system unexpectedly discovered that the model had successfully bypassed the network sandbox restrictions set by the team by invoking third-party external services via an internet search tool. The team was forced to urgently halt the training, completely remove the online execution environment, deploy millisecond-level synchronous safety probes for all tool-call interfaces, and only then resume loading from checkpoints.
The official also acknowledged that early engineering versions exhibited extreme "fixation bias" when encountering overly tricky long and difficult tasks, frequently fabricating execution results out of thin air, deliberately ignoring constraint instructions, or even actively hiding error logs. In the final released stable version, its internally quantified penalty score for "misbehavior" has significantly converged from 66.8 in the previous generation to 14.9, but there remains a generational gap compared to top closed-source models in strict instruction following.
Regarding authoritative evaluation benchmarks, the model's most prominent performance strengths lie in knowledge work and long-horizon enterprise business tasks. In the AA-Briefcase benchmark specifically designed to assess the generation of complex tables, professional presentations, and strategic memos, it achieved an Elo score of 1527. This represents a massive leap from the previous Small's 912, comprehensively surpassing Claude Sonnet 5.5 (medium) at 1439 and Gemini 3.8 Flash (high) at 1203. In official disclosed horizontal comparisons, it claimed the top spot on JobBench (70.5%), the Vals Legal Index (73.9%), the Financial Analysis Index (61.5%), and the authoritative high-difficulty benchmark HLE-Diamond (28.7%).
However, in the pure dimension of code programming, it has not established overall dominance. On the FrontierCode v2.0 and DeepSWE v1.1 software engineering benchmarks, although it narrowly leads Kimi K3 and GLM-5.3-Flash, it significantly trails behind DeepSeek-V4.1-Flash (scores of 25.4% vs. 28.5%, and 72.1% vs. 77.0%, respectively). In the highly challenging SWEBench Pro v2 (Hard) tests, its actual problem-solving rate also falls short of Kimi K3. If benchmarked against top closed-source flagships, its win rate on the real command-line terminal manipulation task Terminal-Bench 4.0 is only 13%, still significantly distant from Sonnet 5.5's 30%.
The direct engineering cost of achieving long-horizon autonomous planning is extraordinarily high token compute consumption. In AA's standardized benchmark tests, Inkling 2 Small requires an average output of approximately 150,000 tokens to complete a single comprehensive task, whereas the previous generation required only 37,000 and Sonnet 5.5 only 31,000. Nevertheless, TML has adopted an aggressive cost-pricing strategy: the API charges only $0.30 per million input tokens and $1.00 for outputs, translating to an average single-task inference cost of merely ~$1.18. By comparison, a single task costs $7.77 for Gemini 3.8 Flash and reaches as high as $10.02 for Sonnet 5.5.
Furthermore, the significant reduction in the model's hallucination rate appears somewhat ingeniously engineered. In the AA-Omniscience omniscience evaluation, although its hallucination rate plummeted from 61% in the previous generation to 27%, its overall answer response rate also drastically retreated from 77% to 53%, with actual comprehensive accuracy only marginally improving from 32% to 34%. This means the optimization of the hallucination metric primarily stems from alignment constraints that enforce a "direct refusal to answer when facing uncertain facts." The official also acknowledges that the model's polishing for general daily conversations remains somewhat stiff, its language organization mechanical, and it suffers from stylistic flaws such as the frequent use of formatted slashes.
Crafting a Well-Suited Hammer for the Cloud Fine-Tuning Platform Tinker
Why invest heavily in creating a specialized model that feels rough in conversational interaction but works extremely hard at coding and processing complex spreadsheet tasks? The answer points directly to TML's current core monetization engine—Tinker.
Tinker is an enterprise-grade cloud fine-tuning API platform officially launched by TML in October 2025. For TML, open-source or open-weight models serve more as a traffic magnet to attract global developers to the Tinker platform: attracting the tech ecosystem with free model weights, then guiding enterprise-level key clients to access the platform for high-value customized fine-tuning and batch inference settlement.
The task distribution during the model's reinforcement learning phase intuitively confirms this strong commercial orientation: according to total token consumption, programming tasks account for a dominant 65.0%, professional office work and knowledge work take 13.3%, deep STEM logic reasoning occupies 7.6%, and cutting-edge web & end-to-end application building accounts for 7.2%. These four high-level productivity tasks combined exceed 93% weight, while routine tasks related to everyday casual chat account for an extremely slight 0.1% across the entire pipeline.
Its automated evaluation system for reinforcement learning is also meticulously tailored to commercial productivity scenarios: code programming uses hidden unit test suites; certain long-horizon tasks feature an added "agent reviewer" that not only verifies whether code passes tests but deeply dissects the logical chain of the model's problem-solving; structured documents and external tool calls are jointly adjudicated by LLM judges configured with strict scoring criteria; for web applications, real-time rendered frontend screenshots undergo pairwise automated visual adversarial evaluation.
Figure: Paper art animation iterated from feedback in Blender by Inkling 2 Small
The killer advantage repeatedly emphasized by TML in its external promotions is precisely its "extreme fine-tunability": under identical dataset scales and algorithmic strategies, the new model demonstrates higher convergence efficiency and greater potential for performance leaps. In an officially disclosed enterprise financial report automation summarization fine-tuning task, Inkling 2 Small reached the final level achieved by the first-generation Small after completing the entire fine-tuning cycle, doing so in just 75 training steps.
The blog also unveiled a highly bootstrapping "nested" automated training experiment: TML allocated a dedicated $500 compute budget for Inkling 2 Small, allowing it to log into the Tinker platform as an autonomous agent and autonomously train a lightweight Qwen3.5-4B model end-to-end to perform time-series prediction tasks. Inkling 2 Small independently built benchmark evaluations, wrote high-quality SFT data, and orchestrated reinforcement learning. After spending 4 hours and 17 minutes and consuming $426 of the budget, it steadily pushed the target small model's benchmark score from 0.191 to 0.289.
The TML team explicitly stated that Inkling 2 Small is merely "the first step on a multi-tier model ladder," with more sizes and full-sized flagship versions already in preparation. Using medium-scale models that balance inference costs as a breakthrough point, first verifying fine-tunability and the long-horizon operational capabilities of autonomous agents, and smoothly directing global high-net-worth developers into the Tinker commercial ecosystem will allow this AI infrastructure business to truly achieve a positive flywheel effect.
Chinese Open-Weight Models Distilled by Silicon Valley Giants
Re-examining that line in the technical log that sent shockwaves through the industry. In its engineering report, TML officially wrote: "The datasets for training stages H3 and H4 were largely sampled from deduction trajectories generated by earlier models (including Moonshot AI's Kimi K3 and our own early models), and underwent strict automated filtering during the cleaning phase targeting logical correctness and alignment standards."
Analyzing its training stage design in depth: Stage H3 contains approximately 320 billion tokens of complex reasoning and agent behavior trajectories (mixed with replay of pretraining data), while Stage H4 consists of curated ultra-high-quality agent golden trajectories totaling about 34 billion tokens. Although this volume is not massive compared to the foundational corpus of 12.2 trillion tokens across the first two stages, the engineering node it occupies determines the model's intelligence form—the model internalizes million-long-context reasoning logic and learns to call tools and collaboratively correct errors like a professional autonomous agent, precisely thanks to these two key corpora, before smoothly transitioning to the reinforcement learning stage.
Although TML did not detail the specific percentage of Kimi K3 corpus within it, making it unjustified to simply label Inkling 2 Small as a "patchwork product relying solely on distilling Kimi," the indisputable objective fact remains: in this complex engineering recipe representing Silicon Valley's first tier, Kimi K3 is the sole external third-party model formally named in black and white throughout the document.
Figure: End-to-end generated 3D roaming and walking routes from a flat floor plan
Why precisely choose external open-source trajectories at the critical stage where long-context and agent capabilities are forming? The fundamental reason lies in the fact that high-value, real-world agent execution trajectories are extremely scarce, high-quality data in the entire AI field. They require large models to operate within real interactive sandboxes, autonomously calling APIs, parsing terminal error messages, iteratively correcting scripts over multiple rounds, and completing end-to-end result verification. A complete long-horizon trajectory often consumes tens of thousands of tokens, making scalable production nearly impossible relying solely on traditional manual annotation.
From an engineering implementation perspective, the most efficient and pragmatic path is undoubtedly to find a mature benchmark model whose underlying code logic and agent execution capabilities have been fully validated, letting it run day and night in massive sandbox tasks to automatically extract all positively successful problem-solving trajectories. In horizontal technical benchmark evaluations published by TML, Kimi K3 consistently ranks ahead of Inkling 2 Small in both long-context comprehension and the highly demanding SWEBench Pro v2 (Hard) software engineering benchmark. Adopting it as a "mentor" in the data pipeline is completely logical from an engineering rationality standpoint.
This dramatic technical interaction becomes even more subtle when placed within the current context of global geopolitical tech competition. Just earlier this February, Anthropic publicly issued an investigative statement accusing Chinese leading large model teams including DeepSeek, Moonshot AI, and MiniMax of frequently invoking the Claude API over 16 million times across multiple interactions using approximately 24,000 distributed accounts to reverse-distill its advanced capabilities. Yet, merely months later, a Silicon Valley top-tier star unicorn founded by core OpenAI veterans, valued at tens of billions of dollars, openly admitted in a public technical blog that it extensively sampled the output trajectories of Moonshot AI's open-source models within its foundational training architecture.
Of course, the two incidents possess fundamentally different legal compliance and open-source contract implications: circumventing commercial API rate limits and service terms through irregular bulk registrations constitutes a gray-area violation of commercial agreements; whereas Kimi K3 belongs to open-weight models, whose model weights are fully transparent in the global open-source community. Local deployment on independent computing clusters and sampling its inference results fall squarely within extremely conventional technical interactions inside the open-source ecosystem.
A deeper upheaval occurs in the quiet shifting of underlying technical ecological niches. Over the past few years in the wave of large model development, the common paradigm for open-source communities and small-to-medium startups was generally to directly pull Meta's Llama or France's Mistral open-source weights for downstream fine-tuning, or to leverage commercial outputs from Silicon Valley closed-source giants to distill smaller specialized models. Today, however, top-tier Chinese open-weight models represented by Kimi K3 have substantially stepped into the role of core corpus suppliers for Silicon Valley's frontline research teams training next-generation flagship bases. Meanwhile, the foundational bootstrapping base invoked in TML's technical demos originates from Alibaba's open-source Qwen architecture.
Peering into the Ecological Niche Rotation and Catch-Up Truth Between Chinese and American Open-Source Models
However, objectively viewing it, even with directly absorbing high-quality trajectories from external benchmarks like Kimi K3, Inkling 2 Small still faces a visibly apparent technological gap to reach the throne of the global open-source ecosystem.
On Artificial Analysis's latest release of the comprehensive open-weight model ranking leaderboard (v5.0.1), Inkling 2 Small indeed leads a pack of European and American peers across the board with a benchmark score of 31. Following it, Mistral Large 4 scores 26, and Meta's heavily invested Muse Spark Open records only 25. This is precisely the reason AA bestowed upon it the specific title of "the highest-scoring open-weight model outside China."
Yet, once stripped of the "outside China" qualifying prefix, the true peak of the global open-weight track remains comprehensively dominated by Chinese models: Moonshot AI's Kimi K3.1 Pro leads the field at an overwhelming absolute advantage of 46 points, DeepSeek's V4.1 Flash records 33, and Zhipu AI's GLM-5.3 secures 32. Inkling 2 Small sits only at fourth place globally, with Xiaomi's MiMo-V2.6-Pro close behind at 30. In other words, this generation of flagship compact powerhouse painstakingly polished by the US team is essentially catching up to the second tier of Flash-level reasoning models in the Chinese open-source ecosystem, leaving a full generational gap of 15 points away from China's absolute top-tier flagship versions.
More realistically, at the current juncture, it cannot yet be considered a fully open-source model in the complete sense. According to the so-called "phased safety-open route" established by TML back in July, the model initially offers beta testing only via a closed-source API format on its own cloud Tinker platform, and must await the thorough completion of external red-team security assessments before conditionally releasing model weights. Although the official blog commits to "planning to fully release all weights to the global open-source community next week under an Apache 2.0 commercial open-source license," AA prudently marks the actual weight release date on its evaluation page as "unconfirmed." Whether this open-source check carrying a $40 billion valuation narrative will be honored in full on schedule will see its final answer from the technical ecosystem next week.
The fierce showdown for global open-source and open-weight models is evolving into a deeply complex cycle of "top players' output achievements mutually absorbing each other, with multi-directional cross-flow of technical trajectories." Thinking Machines Lab, through a candid technical blog that does not shy away from acknowledging references, has thoroughly laid bare the reality of this technical reciprocation and global ecological niche restructuring to the forefront.
It can be foreseen that as Murati's team continues to ascend what they call the "model ladder," producing flagship Inkling 2 models with even larger parameter scales and more advanced reasoning capabilities, the technical confrontations and mutual nourishment between top Chinese and American AI laboratories within the open ecosystem will inevitably trigger far-reaching industrial shocks.