Thinking Machines releases Inkling, strongest U.S. open-weight model to date

Thinking Machines Lab announced Inkling, its first fully released open-weights foundation model family, on July 14–15, 2026. The flagship model, called Inkling, is a Mixture-of-Experts transformer with 975B total parameters and 41B active parameters per token, supporting a 1M token context window on open weights. It was trained from scratch on 45 trillion tokens of text, images, audio, and video. Alongside Inkling, the company previewed Inkling-Small, a lighter-weight variant with 276B total parameters and 12B active parameters, trained with a similar recipe.

The company positioned Inkling as a customizable multimodal base model rather than a benchmark-maxed flagship. Mira Murati described it as the company's "first model" and "trained from scratch" with open weights and same-day fine-tuning on the Tinker platform. Soumith Chintala emphasized open weights, the 975B parameter count, native multimodality, and availability on Tinker, Hugging Face, and launch partners. John Schulman provided timeline context: pretraining began last winter, and from mid-January a small team built coding, reasoning, and agentic training on top. Lilian Weng characterized it as a foundation model aimed at "solid performance across a broad categories of capabilities" and intended for practical use plus customization. TML staff repeatedly stressed that this is a day-1 release and a foundation for future iterations rather than a final frontier push.

Inkling's architecture incorporates several distinctive technical choices. It employs hybrid and sliding-window attention with a 5:1 local-to-global layer ratio and a window size of 512, relative positional encoding (relative attention bias instead of RoPE, which commentators flagged as unusually novel at large scale), short convolution layers around attention and feed-forward streams (unusually scaled-up usage in the community's assessment), and a Mixture-of-Experts design with shared expert sinks and 2 shared experts (atypical since many recent MoEs use 1). The model uses DeepSeek-style auxiliary-loss-free load balancing and incorporates muP and Muon/weight decay variants (Aaron Defazio confirmed the use of his corrected weight decay approach, MuonC/AdamC). For inference acceleration, it includes 8 MTP heads for speculative decoding.

On benchmark performance, Inkling received mixed assessments. Artificial Analysis ranked it at 41 on the Intelligence Index, making it the leading U.S. open-weights release, ahead of Nemotron 3 Ultra (38), Gemma 4 31B (29), and gpt-oss-120b (24). It also scored a GDPval-AA v2 Elo of 1238, higher than Kimi K2.6 (1190) and DeepSeek v4 Flash max (1189), and achieved 24% on τ³-Banking, above Kimi K2.6 (21%) and slightly above DeepSeek v4 Flash max (23%). On agentic tasks, Design Arena reported Inkling entered the Agentic Web App Arena at #9 overall with an Elo of 1257, in the same band as Claude Opus 4.6 and Gemini 3.5 Flash, and called it the highest-ranking U.S.-based open-weight model for agentic workloads. Artificial Analysis also noted that Inkling averages 25K output tokens per Intelligence Index task, positioning it as relatively token-efficient compared to GLM-5.2 max (43K), Kimi K2.6 (38K), and DeepSeek v4 Pro max (37K). However, Natolambert called it a "clear step up from Nemotron Ultra" and "new best American model" but still "a bit behind GLM 5.2 on agentic benchies, and Kimi K 2.6 on multi modal." Scaling01 argued the benchmarks are "not that great," describing it as roughly "another Kimi-K2.6" and behind all closed models and GLM-5.2, speculating the release may have been timed ahead of Kimi-K3 and DeepSeek-V4-GA. Stochasticchasm said it seems "very strong for multimodal" but "not super strong for terminal bench etc." JJitsev pushed back on hype around Inkling being the "only open-weight model trained without distilling," noting it uses distillation from open weights and underperforms GLM 5.2 on TerminalBench-style evals; however, TeortaxesTex offered a contrarian spin, suggesting that mediocre benchmark-maxing may actually indicate less corner-cutting and distillation contamination and a more independent data pipeline. Alex Kirillov claimed Inkling avoids the common "audio in = intelligence penalty" seen in many omni models, though another user asked for stronger supporting evidence and benchmarks.

On inference and optimization, ecosystem partners delivered substantial day-0 support. NVIDIA reported that Inkling was trained on GB300 NVL72 and that an NVFP4 checkpoint was available on Hugging Face on day 0. vLLM said day-0 support includes NVFP4 and BF16 optimizations, reaching up to 380 tokens per second per user on 4× GB200 with MTP. Inferact detailed sconv-aware tensor-parallel sharding, low-latency fused collectives (5× faster at batch size 1), and direct integration of TML's FA4 sheared-bias kernel. LMSYS/SGLang implemented native Inkling architecture support including ShortConv, relative positional attention, shared expert sink MoE, prefill full CUDA graph, MXFP8 KV cache, full parameter and LoRA RL in a customized Megatron backend, routing replay, cross-runtime parameter sync, and DFlash speculative decoding. Modal provided a custom DFlash speculator yielding 67% higher throughput and interactivity, with Soumith Chintala amplifying that Modal's DFlash speculator is "much faster than MTP." Lysandre reported that replacing TML's causal Conv1D with causal-conv1d yielded +4% throughput, and replacing attention with FlashAttention-4 yielded another +11%, for approximately 15% total throughput gain without retraining. Unsloth released 1-bit GGUF quantizations claimed to be 86% smaller (270GB vs 1.9TB) while retaining 74.2% of top-1% accuracy, with vision and audio support.

On pricing and availability, Artificial Analysis listed Tinker API pricing as $1.87 per 1M input tokens and $0.374 cached for the 64K context tier, and $3.74 per 1M input and $0.748 cached for 256K context, with output pricing of $4.68 and $9.36 respectively. Full open weights were released on Hugging Face, and the model is available via Tinker, Databricks, Baseten, Modal, and vLLM/SGLang serving stacks. The model carries an Apache 2.0 license, permitting wide commercial and research use. Observers including Natolambert, Karin Anguyen, and others celebrated the open-weight and permissive licensing as a major boost to the U.S./Western open ecosystem. Researchers and builders praised the explicit framing that Inkling is a broad, tunable foundation rather than a benchmark-maxed point solution, and several users highlighted the transparency, grounded tone, and comprehensive technical documentation.

Thinking Machines' release of Inkling marks a deliberate pivot in open-source AI strategy away from benchmark-chasing toward a customizable foundation model positioned for practical deployment and fine-tuning. The company explicitly framed Inkling as a day-1 foundation for future iterations rather than a final frontier push, a messaging choice that distinguishes it from the typical "SOTA claim" playbook. The Apache 2.0 license and immediate availability across seven major serving platforms (vLLM, SGLang, Modal, Baseten, Databricks, Hugging Face, and NVIDIA) signal a deep commitment to ecosystem integration—a meaningful contrast to models that ship narrowly or with restrictive terms.

Inkling's technical architecture reflects several unconventional choices that attracted attention from the research community. The use of relative positional encoding instead of RoPE, hybrid sliding-window attention with a 5:1 local-to-global ratio, short convolution layers around attention/FFN streams, and DeepSeek-style auxiliary-loss-free load balancing with shared expert sinks all depart from recent defaults. These choices may indicate either independent architectural exploration or deliberate avoidance of standard recipes—a point of debate among observers. Pretraining began last winter, with coding, reasoning, and agentic training layered on from mid-January onward, suggesting a relatively compressed timeline to general release.

On performance, Inkling's standing is mixed. Artificial Analysis ranked it as the strongest U.S. open-weight model (Intelligence Index 41, vs. Nemotron 3 Ultra at 38) and it entered Agentic Web App Arena at #9 overall with an Elo of 1257, placing it in the same band as Claude Opus 4.6 and Gemini 3.5 Flash. However, independent commentators flagged gaps: Natolambert called it "a bit behind GLM 5.2 on agentic benchies, and Kimi K 2.6 on multi modal," and Scaling01 described it as "roughly another Kimi-K2.6" and behind all closed models and GLM-5.2. The contrast between strong U.S. rankings and mixed global performance reflects the current landscape, in which Chinese open models have surged ahead. Some observers framed Inkling's moderate benchmark scores as evidence of less corner-cutting and distillation contamination; others saw it as a strategic timing choice ahead of announcements from competitors.

No comments yet. Be the first to share your thoughts!