The open-weights LLM landscape just gained a significant new entrant. Inkling, released on July 15 by Thinking Machines Lab, is a 975B-parameter Mixture-of-Experts model with 41B active parameters — and it arrives with a refreshingly different pitch: it’s not trying to top every leaderboard. Instead, it’s designed as a multimodal foundation model optimized for fine-tuning and customization.
That positioning is worth paying attention to. Most model releases frame themselves as “the best yet” on a curated benchmark suite. Inkling’s announcement explicitly states it is “not the strongest overall model available today, open or closed.” What it offers instead is a combination rarely found in a single open-weights release: native text/image/audio multimodality, controllable thinking effort, strong safety properties, and full availability for fine-tuning on the company’s Tinker platform — all under an Apache 2.0 license.
Let’s break down the architecture, training methodology, and the design decisions that make this release interesting.
Architecture: Familiar MoE with Notable Departures
Inkling is a 66-layer decoder-only transformer with a sparse MoE feed-forward backbone. The routing setup follows the DeepSeek-V3 playbook: each MoE layer has 256 routed experts with 6 active per token, plus 2 shared experts active on every token. A sigmoid-based router with auxiliary-loss-free load-balancing bias handles expert selection.
Two architectural choices stand out as departures from the current norm:
- Relative positional embeddings instead of RoPE. While rotary positional embeddings (RoPE) have become the de facto standard for modern transformers, Inkling uses relative position representations — the approach described by Shaw et al. (2018) and later used in Music Transformer. The team reports it performs better and extrapolates more reliably to longer sequences than RoPE, which matters given the 1M-token context window.
- Short convolutions on residual branches. Inkling applies short 1D convolutions at two points: after the key and value projections in each attention layer, and on the attention and MLP residual branch outputs before they rejoin the main residual stream. This is a lightweight addition that can improve local feature extraction without significant overhead.
For attention, Inkling interleaves sliding-window and global layers at a 5:1 ratio with 8 KV heads. The hybrid approach balances long-range context access with computational efficiency — global layers handle full attention while the more frequent sliding-window layers keep per-token cost down.
Encoder-Free Multimodality
Many multimodal models bolt on separate vision encoders (like CLIP) and audio encoders. Inkling takes a different route: an encoder-free architecture where audio and vision inputs are tokenized directly and processed jointly with text through the same decoder.
- Audio: Input as dMel spectrograms — discrete speech tokens that eliminate the need for a separate audio encoder. Inkling transcribes speech, follows spoken instructions, and reasons over longer recordings. On VoiceBench it scores 91.4%, and on MMAU (a multi-turn audio understanding benchmark) it hits 77.2%.
- Vision: Images are encoded as patches of 40×40 pixels using a four-layer hMLP. Inkling demonstrates strong performance on charts, diagrams, and mathematical visual reasoning — scoring 73.5% on MMMU Pro (Standard 10), which is competitive among open-weights models, though it trails closed-weights models like Gemini 3.1 Pro (82.0%) and Claude Fable 5 (84.2%).
The encoder-free design was chosen for consistency with Thinking Machines’ interaction model system — a real-time collaborative interface that uses voice and vision naturally. It also means the model is simpler to deploy: no separate encoder components to manage.
Controllable Thinking Effort
Perhaps the most practically useful feature for developers is Inkling’s controllable thinking effort. The model accepts an effort parameter from 0.2 to 0.99 that adjusts how many reasoning tokens it generates before producing its answer. This is not just a truncation knob — the model was trained to modulate its reasoning depth based on this setting.
The efficiency gains are substantial. On Terminal Bench 2.1 (an agentic coding benchmark), Inkling at a given effort level achieves the same score as Nemotron 3 Ultra using roughly one-third the tokens. For workloads where you’re running a model millions of times — grading, synthetic data generation, agentic loops — that token efficiency directly translates to cost and latency savings.
This matters because cost and latency are often the binding constraints in production, not peak benchmark performance. A model that can match a competitor at one-third the token budget gives developers meaningful room to tune the cost-performance tradeoff per use case.
Training: Muon Optimization and 30M+ RL Rollouts
Inkling was pretrained on 45 trillion tokens spanning text, images, audio, and video, using a hybrid optimization strategy: Muon for large matrix weights and Adam for other parameters. The team coupled weight decay strength to the square of the learning rate, which kept model weight magnitudes stable across long training horizons. Training ran on NVIDIA GB300 NVL72 systems.
The post-training pipeline is where things get interesting. After an initial SFT phase bootstrapped on synthetic data from open-weights models (including Kimi K2.5), Thinking Machines ran large-scale asynchronous RL with over 30 million rollouts. Reasoning performance improved log-linearly throughout the entire process — no plateau, no instability across two long continuous training runs.
One emergent behavior stood out: the model’s chain-of-thought became progressively more compressed over the course of RL training. It dropped grammatical overhead — articles, connectives, full sentences — in favor of terse, telegraphic notation, while remaining comprehensible and reaching the same answers. Crucially, this compression was never targeted by any reward signal. It emerged purely from the efficiency incentives of the RL loop. The same effect was recently noted by the Cognition team during training of SWE-1.7, suggesting it may be a general property of large-scale reasoning RL.
Calibration and Forecasting
Most model releases focus on accuracy. Inkling puts unusual emphasis on calibration — the model’s ability to express appropriate confidence in its answers, including knowing when to say “I don’t know.” This was trained directly: RL against proper scoring rules on a large corpus of resolved real-world questions, plus short-form factual QA with abstention-aware rewards that only pay off when the model is likely to be right.
The result shows up in forecasting benchmarks. On ForecastBench (Brier Index, no search), Inkling scores 61.1 — matching Gemini 3.1 Pro and outperforming GPT-5.5 (59.1) and Claude Opus 4.8 (54.6). On Prophet Arena (Brier Score, lower is better), Inkling achieves 0.1617, competitive with the best models tested. For developers building systems where calibrated uncertainty matters — decision support, risk analysis, prediction markets — this is a meaningful differentiator.
Safety: Best-in-Class Among Open Weights
On FORTRESS — a benchmark measuring refusal of harmful requests alongside benign look-alikes — Inkling scores 78.0% on the adversarial track, the highest among all open-weights models tested (vs. Nemotron 3 Ultra at 77.6%, GLM 5.2 at 71.3%, and DeepSeek V4 Pro at 36.0%). Equally important, its FORTRESS Benign score is 95.9%, meaning it avoids over-refusing safe requests. On StrongREJECT, it scores 98.6%, in line with both open and closed-weights models.
Inkling-Small Preview
Alongside the main model, Thinking Machines shared a preview of Inkling-Small: a 276B-parameter MoE model with only 12B active parameters. Despite being roughly one-third the active parameter count, Inkling-Small matches or exceeds its larger sibling on several benchmarks — 87.2% → 88.3% on GPQA Diamond, 79.8% → 83.4% on IFBench, 78.1% → 76.7% on Charxiv RQ (where Inkling wins). On SWE-Bench Verified, both land at ~77.5%.
The smaller model is positioned for cost-sensitive workloads: coding assistance, LLM-based grading, and synthetic data generation. Full weights will be released once testing is complete.
Deployment and Ecosystem
Inkling’s full weights are available on HuggingFace in both BF16 and NVFP4 formats (the latter optimized for NVIDIA Blackwell inference). The model is supported across a broad deployment ecosystem on day one:
- Local inference: SGLang, vLLM, llama.cpp via Unsloth, HuggingFace transformers
- API providers: Together AI, Fireworks AI, Modal, Databricks, Baseten
- Fine-tuning: Available on Tinker with 64K and 256K context options, plus an interactive playground for testing the model before committing to a training run
The Tinker Cookbook on GitHub includes recipes for fine-tuning Inkling, including examples that leverage its unique audio capabilities.
What This Means for the Open-Weights Landscape
Inkling doesn’t win on raw benchmark numbers — closed-weights models like GPT 5.6 Sol and Claude Fable 5 outperform it on most reasoning and coding evals. But that’s not the point. What Inkling offers is something the open-weights ecosystem has been missing: a genuinely multimodal model (text, image, and audio in one architecture), with controllable inference cost, strong safety properties, and a clear path to customization.
The controllable thinking effort feature alone could change how teams architect their AI pipelines. Instead of choosing between a cheap model and a smart model, you get a single model that can be both — adjusting its reasoning budget per request based on the task’s difficulty. Combined with Apache 2.0 licensing and broad inference support, Inkling is a strong candidate as a base model for organizations that need to fine-tune rather than just prompt.
The open-weights race in 2026 is no longer just about who has the highest score. It’s about which model gives developers the most control — over cost, over behavior, over customization. By that measure, Inkling is a serious contender.