Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati in February 2025, has released its second-generation model, Inkling 2 Small. The Mixture-of-Experts model carries 594 billion total parameters with 17 billion active during inference, supports image inputs, and is being offered initially through the company's Tinker platform with closed weights. The company says it intends to make the full weights available under an Apache 2.0 license next week after the completion of external safety testing and red-teaming [1][5].
The launch arrives weeks after Business Insider reported the company is nearing a $5 billion funding round led by Accel at a $40 billion pre-money valuation—more than three times the $12 billion valuation attached to its $2 billion seed round in July 2025 [2]. Thinking Machines declined to comment on the financing; Accel did not respond to requests for comment [2].
What the model claims to do #
According to the company's blog post, Inkling 2 Small "far surpasses our previous models' scores on the Artificial Analysis Intelligence Index with a score of 31, reflecting gains in performance on agentic coding, vision, knowledge work, and reasoning" [1][5]. The company attributes the improvement to "more efficient pre-training and a large, unified RL run" spanning hundreds of thousands of long-horizon coding tasks across multiple agent harnesses [1][5].
Third-party evaluator Artificial Analysis reported an Intelligence Index score of 31 (the company's model card lists 30.6 on v5.0)—an 18-point jump over the prior Inkling Small—but found the model generates roughly four times as many output tokens per task, raising the average cost to $1.18 versus $0.20 for its predecessor [2]. On AutomationBench-AA, which tests automated business workflows, the score rose from 5% to 46%; on FrontierCode 2.0 for software engineering, from 5% to 25% [2].
Benchmark performance: mixed against peers #
The company publishes extensive head-to-head tables comparing Inkling 2 Small against open-weight models GLM-5.3-Flash, DeepSeek-V4.1-Flash-1017, Muse Spark Open, and Kimi K3, plus closed models including GPT-6.1 Sol, Gemini 3.8 Flash, and Claude Haiku 5.5 [5][6]. The full comparison appears in the model card [6].
On coding benchmarks, Inkling 2 Small ranks second behind DeepSeek-V4.1-Flash-1017 on two of six reported benchmarks: FrontierCode 2.0 (25.4% vs 28.5%) and DeepSWE v1.1 (72.1% vs 77.0%) [5][6]. On SWEBench Pro v2 Hard, Inkling 2 Small scores 86.3%, second behind Kimi K3 at 88.2%; DeepSeek-V4.1-Flash-1017 does not report a score on this benchmark [5][6]. On Vibe Code Bench v1.1, it places third behind DeepSeek (87.7%) and Kimi K3 (85.0%); on PostTrainBench v1.2, it ranks second behind GLM-5.3-Flash (36.8%); and on Terminal Bench Science v0.1, it trails both DeepSeek (12.9%) and Kimi K3 (8.1%) [5][6].
Inkling 2 Small leads on several visual reasoning tasks (DesignArena Website: 1340, highest reported; GDP.pdf: 25.2%; PerceptionBench: 64.7%) and finance benchmarks (Vals Finance Index: 61.5%; DAYJOB Finance: 14.3%) [5][6]. On the Artificial Analysis Intelligence Index v5.0, it scores 30.6, behind DeepSeek-V4.1-Flash-1017 (32.6) and Claude Haiku 5.5 (36.5) [6]. The company's cited score of 31 differs slightly from the third-party index version reported in the model card (30.6 on v5.0), possibly due to index version or rounding [1][5][6].
The company also highlights an autoresearch demonstration: Inkling 2 Small autonomously fine-tuned Qwen3.5-4B on time-series forecasting, substantially improving the base model's skill in 4.5 hours at $426 compute cost [5]. The reported skill metric rose from 0.191 (base model with proper seeds) to 0.289 (final model), an increase of about 50%; the company's blog says the model "more than doubled its forecasting score." Such demos are self-reported and not independently replicated.
Training and infrastructure #
Pretraining processed 8.4 trillion tokens of text, code, and image data in 68 hours on Nvidia GB300 NVL72 systems using the NVFP4 4-bit format native to Blackwell GPUs, followed by 3.8 trillion tokens of newer data at 512K context [2][5]. A joint reinforcement learning run—combining reasoning and agentic tasks in a single pipeline, unlike the separate runs used for the first Inkling—spanned 1,075 steps and over 17 million rollouts across nine days on 2,048 GPUs [2][5]. The company says it uses sticky routing and DRAM KV-cache off to maintain 93%+ prefix cache hit rates at long contexts [5].
Funding and valuation context #
The July 2025 seed round included Nvidia, AMD, Cisco, and Jane Street [1]. In March 2026, the company announced a partnership with Nvidia to deploy at least one gigawatt of Vera Rubin systems beginning in 2027, alongside an undisclosed Nvidia investment [2]. The reported $5 billion Accel-led round at a $40 billion pre-money valuation would bring post-money to roughly $45 billion—nearly quadrupling the valuation in 16 months (approximately 3.75x the seed post-money)—though neither party has confirmed [2]. The company has disclosed little about revenue or financial performance [2].
Availability and open-weight timeline #
Inkling 2 Small is accessible now via the Tinker Playground for inference and through Tinker for fine-tuning and sampling [2][5]. Weights remain closed for now. The company states: "We intend to make the full weights available under an Apache 2.0 license next week after the completion of external safety testing and red-teaming" [1][5]. That timeline is a company statement, not a guaranteed date.
Acknowledged limitations #
The company's own blog lists several deficiencies: the model "lacks full optimization for conversational interfaces, exhibiting deficiencies in textual formatting, excessive forward-slash generation, and heightened sensitivity regarding mental health and user safety requests," and it "displays verbose phrasing and occasionally misallocates tool calls when handling software environments with extensive toolsets" [1][5]. Safety evaluations show the model trails closed-weight frontiers on misbehavior (14.9 vs 0.8 for GPT-6.1 Sol) and covert misbehavior (11.8 vs 0.8) but leads on destructive actions (4.7 vs 7.0 for GPT-6.1 Sol), while improving substantially over prior Inkling versions [5].
What remains unknown #
- Independent replication of the autoresearch and agentic coding claims.
- Whether the reported $5 billion funding round closes on the reported terms.
- Real-world deployment costs at scale, given the 4x token multiplier noted by Artificial Analysis.
- How the model performs on tasks outside the published benchmark suites.
- The exact composition and provenance of the 12.2 trillion tokens used in pretraining, described only as "publicly available sources, acquired from third-parties, generated internally, or synthetically generated or augmented" [6] .
Thinking Machines Lab has shipped a technically ambitious model with credible infrastructure behind it. The benchmark gains are real within the reported suites, and the open-weight commitment—if fulfilled on schedule—would make it the highest-performing open model from a non-Chinese company by several third-party indices. But the funding narrative, the cost structure, and the gap to closed-weight frontiers on alignment and reasoning remain open questions.
Sources #
- Thinking Machines unveils Inkling 2 Small model with 594 ... (bignewsnetwork.com)
2. [Thinking Machines Launches Inkling 2 Small as Reported $5B ...](https://techstrong.ai/articles/thinking-machines-launches-inkling-2-small) (techstrong.ai)
3. [Inkling 2 Small - Hacker News](https://news.ycombinator.com/item?id=50386338) (news.ycombinator.com)
4. [The AI company founded by the former CTO of OpenAI has ...](https://gigazine.net/gsc_news/en/20261106-inkling-2-small) (gigazine.net)
5. [Introducing Inkling 2 Small - Thinking Machines Lab](https://thinkingmachines.ai/news/inkling-2-small) (thinkingmachines.ai)
6. [Inkling 2 Small Model Card - Thinking Machines Lab](https://thinkingmachines.ai/model-card/inkling-2-small) (thinkingmachines.ai)
7. [GitHub - thinking-machines-lab/tinker-cookbook: Post-training with Tinker · GitHub](https://github.com/thinking-machines-lab/tinker-cookbook) (github.com)