Thinking Machines Lab launches Inkling open-weights model

Inkling's design is openly acknowledged to draw from DeepSeek V3, but calling it a copy would be inaccurate ^[5]. The underlying blueprint is a 66-layer decoder-only transformer with 256 routed experts and 2 shared experts per MoE layer - with 6 routed experts activating per token ^[2]. This expert-routing density is what makes the 41B active parameter count possible despite a 975B total footprint. The key departure from prior MoE designs is a sigmoid-based router with auxiliary-loss-free load balancing, meaning the model does not penalize the routing loss during training - a technique DeepSeek V3 pioneered that Thinking Machines adapted ^[2].

Attention is handled via interleaved sliding-window and global attention in a 5:1 ratio, trading raw full-context compute for efficiency at most sequence positions ^[2]. Multimodality is woven in at the token level without a separate encoder tower: audio enters as dMel spectrograms and images as 40x40-pixel patches through a four-layer hMLP encoder that feeds directly into the same token stream ^[2]. Training used a hybrid optimizer - Muon for large matrix weights and Adam for all other parameters - on Nvidia GB300 NVL72 systems ^[3]. The combination of architectural borrowing from a proven Chinese MoE and US-built hardware infrastructure produced a model trained entirely from scratch in under 9 months ^[1].

The most concrete data point in Inkling's launch is the Bridgewater Associates collaboration, which produced a fine-tuned version scoring 84.7% on financial reasoning benchmarks at less than one-tenth the cost of proprietary alternatives ^[7]. This is not a headline number Thinking Machines generated internally - it is a live enterprise deployment by one of the world's largest hedge funds on a task where accuracy failures have real financial consequences. The cost ratio matters more than the accuracy score: if a fine-tuned open-weights model can achieve this level of accuracy at less than one-tenth the cost of proprietary APIs, the calculus for regulated financial institutions shifts decisively toward self-hosting.

This is the business model Holger Mueller called potentially the biggest innovation from Thinking Machines - not the model itself but the pipeline from open weights to Tinker fine-tuning to domain-specific deployment ^[7]. Analysts at Futurum noted the shift from per-token API pricing to infrastructure the enterprise controls as the core economic argument ^[7]. The Apache 2.0 license removes the legal barrier to commercial fine-tuning and redistribution, which proprietary models cannot match. The caveat is infrastructure: BF16 inference requires a minimum of 2TB aggregated VRAM - 8 Nvidia B300 or 16 H200 GPUs - while the NVFP4 quantized version drops that requirement to approximately 600GB ^[2]. For organizations without that hardware, the five inference API partners (TogetherAI, Fireworks, Modal, Databricks, Baseten) provide access without the capital expenditure ^[3].

Open-weights frontier models have until now been dominated by Chinese labs: DeepSeek, Moonshot AI (Kimi), and others. Western enterprises - particularly in finance, defense, and healthcare - face procurement and regulatory friction when deploying models from Chinese-origin organizations, regardless of license permissiveness. Inkling is explicitly positioned as a US-developed, Apache 2.0 alternative that regulated enterprises can self-host without geopolitical exposure ^[6]. Analyst Pareekh Jain framed this directly: Inkling gives organizations a US-developed open-weight option they can deploy on their own infrastructure ^[6]. The immediate community response on X skewed toward ML infrastructure - the vLLM project announced Day 0 support, signaling that the open-source tooling ecosystem treats Inkling as production-grade infrastructure, not a research artifact.

The irony embedded in Inkling's release is that its architecture is derived from DeepSeek V3 and its post-training data was partially generated using Moonshot AI's Kimi K2.5 ^[1]. Thinking Machines acknowledges this openly and notes it plans to bring the full post-training process in-house for its next model ^[1]. The current version therefore occupies a transitional position: US-trained, US-licensed, but architecturally and data-lineage linked to the Chinese open-weights ecosystem it is meant to displace. For enterprises with strict data provenance requirements, this nuance may affect procurement decisions even as the Apache 2.0 license and US origin address the primary compliance concern. On YouTube, early independent reviews framed the launch as a direct test of whether a US lab could credibly compete with Chinese-dominated open-weights rankings - a question the Artificial Analysis benchmark data now at least partially answers.

One of Inkling's less-publicized features is a controllable reasoning effort parameter, adjustable from 0.2 to 0.99 via system message ^[2]. This allows users to set a speed-accuracy tradeoff at inference time without reloading or switching models - useful for agentic pipelines where some subtasks require deep reasoning and others require fast, cheap completions. The practical implication is visible in one benchmark: Inkling averages 25,000 output tokens per agentic task, compared to 37,000-43,000 tokens for competing models ^[4]. Fewer output tokens at equivalent task quality means lower inference cost per completed task, which compounds at enterprise scale.

This efficiency claim is distinct from raw benchmark scores. On the Artificial Analysis Intelligence Index, Inkling scores 41 - 3 points above Nvidia's Nemotron 3 Ultra at 38, which had previously held the top US open-weights position ^[4]. Task-specific benchmarks include AIME 2026 at 97.1%, GPQA Diamond at 87.2%, SWE-Bench Verified at 77.6%, VoiceBench at 91.4%, and MMMU Pro at 73.5% ^[3]. The Tinker API pricing is $1.87-$3.74 per million input tokens and $4.68-$9.36 per million output tokens - meaningfully below frontier proprietary model pricing for organizations that can operate within the 256K context window the Tinker API exposes (versus the 1M context available on the open weights) ^[3].

Alongside the flagship 975B model, Thinking Machines released a preview of Inkling-Small - a companion model with 276B total parameters and 12B active parameters, trained with a similar recipe and targeting low-latency workloads ^[1]. Weights for Inkling-Small are not yet publicly available, making it a preview rather than a full release. The two-model approach mirrors the tiered strategies used by Anthropic (Haiku-Sonnet-Opus), Google (Flash-Pro), and Meta (8B-70B-405B Llama variants): a flagship model that sets capability benchmarks and a smaller model optimized for production deployments where latency and cost per call matter more than peak accuracy.

For Thinking Machines, the small model is strategically important because the infrastructure barrier to self-hosting the full Inkling is significant. At 12B active parameters, Inkling-Small targets the deployment tier where most production agentic workloads actually run - not the benchmark tier where frontier models compete. If Inkling-Small achieves strong performance at that scale, it significantly expands the addressable enterprise market beyond organizations with 2TB VRAM clusters. The timing of the full weight release for Inkling-Small will be a key indicator of how quickly Thinking Machines can convert benchmark interest into deployed production adoption ^[1].