Inkling

Thinking Machines Lab shipped its first open-weight model today, and the pitch is unusual for a lab stacked with ex-OpenAI talent: Inkling isn't trying to top a single leaderboard. It's a 975-billion-parameter mixture-of-experts model that the company itself describes as "not the strongest model available today, closed or open," built instead to be a starting point organizations fine-tune into something narrower and better for their own workloads.

Overview

Mira Murati founded Thinking Machines Lab in February 2025 with a roster that reads like an OpenAI alumni reunion: John Schulman, Barrett Zoph, Lilian Weng, Andrew Tulloch, and Luke Metz all joined as co-founders. The company has moved fast since then, first with the Tinker fine-tuning API, then with a 276B-parameter voice-focused preview model called TML-Interaction-Small, and now with Inkling, its first model shipped with full open weights. TechCrunch's Connie Loizos framed the release as Thinking Machines "amping up its bet against one-size-fits-all AI" - a strategy that puts the company at odds with how Anthropic, OpenAI, and Google package their frontier models as sealed APIs.

Architecturally, Inkling is a 66-layer decoder-only transformer with sparse MoE routing: 256 routed experts plus 2 shared experts per layer, with 6 routed experts active per token. It uses interleaved sliding-window and global attention at a 5:1 ratio, learned relative positional embeddings instead of RoPE, and an encoder-free design that folds text, image, and audio tokens into a single hidden space rather than bolting on separate vision or audio towers. The model trained on 45 trillion tokens spanning text, images, audio, and video, entirely on Nvidia GB300 NVL72 systems - the same compute Thinking Machines locked down through deals with Nvidia and Google Cloud earlier this year.

Post-training leaned on synthetic data bootstrapped from Moonshot AI's Kimi K2.5, followed by reinforcement learning scaled past 30 million rollouts. Thinking Machines says future versions will drop the dependency on another lab's model for that bootstrap step. The company also reports an odd side effect of the RL process: Inkling's chain-of-thought got shorter as training progressed, without losing coherence, which the team attributes to the model finding more efficient reasoning paths rather than being explicitly optimized for brevity.

Key Specifications

A companion model, Inkling-Small, packs 276B total parameters with 12B active and is currently in preview. Thinking Machines says it matches or beats the full-size Inkling on several benchmarks despite the smaller footprint, though full weights are still pending completion of testing.

Mira Murati co-founded Thinking Machines Lab in February 2025 with a team of OpenAI alumni; Inkling is the company's first fully open-weight model release. Source: wikipedia.org

Benchmark Performance

Thinking Machines published head-to-head scores against several closed and open competitors at its "effort=0.99" reasoning setting. Here's how Inkling stacks up against the models we track most closely:

Claude Fable 5 wins every column here, sometimes by a wide margin, which is consistent with Thinking Machines' own framing that Inkling isn't chasing the top of any single leaderboard. The more interesting comparison is efficiency, not accuracy: on Terminal-Bench 2.1, Inkling beats Nemotron 3 Ultra (63.8% vs 56.4%) while Thinking Machines says it uses roughly a third as many tokens per task, a truly useful result for teams billed by output tokens. On Gemini 3.1 Pro's home turf of adversarial safety, Inkling's 78.0% on FORTRESS edges past Nemotron's 77.6% and clears Gemini's 65.2%, though it still falls well short of Fable 5's 96.0%.

One benchmark where Inkling actually leads the field: ForecastBench, a calibration test that scores models on real-world prediction accuracy without search access. Inkling's Brier index of 61.1 (plus or minus 0.79) edges out GPT-5.5 at 59.1 and ties Gemini 3.1 Pro at 61.1, though Grok 4.3 still leads the pack at 61.7. For a model not built to top raw intelligence benchmarks, coming out ahead on calibrated forecasting is a specific, defensible claim rather than a marketing one.

Key Capabilities

Inkling's headline feature isn't a benchmark score, it's the pairing with Tinker. Thinking Machines built the model to be fine-tuned, not just queried, and the company's own demo showed Inkling improving itself on a lipogram-writing task through self-directed fine-tuning on the platform. That's a different pitch than "use our API": customers own the resulting weights and can export them, so a company doing domain-specific work (legal, medical, financial) can shape the model around proprietary data instead of prompting around a frozen one.

The multimodal design is native rather than adapter-based, which shows up in tasks like Charxiv - reading and reasoning over scientific charts - where Inkling improves from 78.1% to 82.0% when given Python execution to verify its own answers. On agentic coding, the model handles long-context refinement tasks (Thinking Machines shows a 40-iteration game-development loop) and ships with embedded browser tool use, putting it in the same agentic category as GPT-5.6 Sol and other tool-calling frontier models, if not at the same accuracy tier.

Bridgewater Associates gave the strongest early third-party validation: a version of Inkling fine-tuned on the firm's own data scored 84.7% on a financial reasoning benchmark, beating proprietary alternatives at roughly one-fourteenth the operating cost, according to Thinking Machines. That's the kind of result the company is betting will matter more to enterprise buyers than a HLE percentage point.

Inkling trained entirely on Nvidia GB300 NVL72 systems, the same hardware generation behind Thinking Machines' compute deals with Nvidia and Google Cloud. Source: pexels.com

Pricing and Availability

Inkling ships with full open weights on Hugging Face, including a quantized NVFP4 checkpoint tuned for Nvidia Blackwell hardware, and runs on SGLang, vLLM, llama.cpp, and TokenSpeed out of the box. Together, Fireworks, Modal, Databricks, and Baseten have already added it to their model catalogs, so teams that don't want to self-host have several inference options at launch.

The primary commercial path, though, runs through Tinker. Pricing there splits into three meters: prefill (input processing), sample (generation), and train (forward and backward passes). At the 64K context tier, Inkling currently runs $1.87/M prefill tokens and $4.68/M sample tokens, both reflecting a limited-time 50% launch discount; training costs $5.61/M tokens. Stepping up to the 256K tier roughly doubles all three rates. Cached prefill tokens get a 80% discount regardless of tier. Tinker also offers free, limited-time playground access to test the model before committing to a paid plan.

That pricing is a starting point, not a fixed rate. Tinker's published schedule shows a broader increase coming July 17, 2026, raising prefill and sample rates by about 50% and training rates by about 10%, which the company attributes to rising compute costs. Anyone budgeting against Inkling for a production fine-tune should check the live pricing page rather than treat launch-day numbers as durable.

Strengths and Weaknesses

Strengths

• Full Apache 2.0 open weights, not a restrictive research license, with day-one support across five inference providers and four open-source runtimes
• Token-efficient agentic performance: beats Nemotron 3 Ultra on Terminal-Bench 2.1 using roughly a third of the tokens
• Native multimodal architecture (text, image, audio) without bolted-on encoders, plus a 1M-token context window
• Leads on ForecastBench calibration, a truly differentiated result rather than a marginal win
• Deep fine-tuning integration through Tinker, with customers retaining ownership of exported weights

Weaknesses

• Trails closed frontier models like Claude Fable 5 by wide margins on HLE, SWE-Bench Verified, and MMMU Pro
• Companion Inkling-Small model isn't fully released yet, weights still pending final testing
• Post-training bootstrap relied on a competitor's model (Kimi K2.5), a dependency Thinking Machines says it wants to shed
• Tinker pricing is scheduled to rise within days of launch, complicating cost planning for new adopters
• No VoiceBench or MMMU Pro numbers reported yet against some direct competitors, leaving gaps in the comparison

Related Coverage

We don't have a hands-on review of Inkling yet. For background on Thinking Machines Lab's infrastructure and prior releases, see our coverage of its Nvidia compute deal, its Google Cloud GB300 partnership, and the earlier TML-Interaction-Small voice model. For where Inkling's coding scores sit against the rest of the field, see the coding benchmarks leaderboard and SWE-Bench coding agent leaderboard.

FAQ

Is Inkling actually open source?

Yes. Thinking Machines released full weights under Apache 2.0 on Hugging Face, including a NVFP4 checkpoint for Nvidia Blackwell, with no restrictions on commercial use or fine-tuning.

How does Inkling compare to Claude Fable 5 or Gemini 3.1 Pro?

Inkling trails both on raw benchmark scores like HLE and SWE-Bench Verified. Thinking Machines positions it as a customizable generalist for fine-tuning, not a leaderboard leader.

What is Inkling-Small?

A companion 276B-parameter model (12B active) in preview that reportedly matches or beats full Inkling on several benchmarks. Full weights haven't shipped yet.

What does Inkling cost to run?

Through Tinker, 64K-context pricing starts at $1.87/M prefill tokens and $4.68/M sample tokens after a limited-time 50% launch discount. Rates are scheduled to rise around July 17, 2026.

What hardware was Inkling trained on?

Completely on Nvidia GB300 NVL72 systems, using 45 trillion tokens of text, image, audio, and video data.

Sources:

• Introducing Inkling - Thinking Machines Lab
• Inkling model card - Thinking Machines Lab
• Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling - TechCrunch
• thinkingmachines/Inkling model card - Hugging Face
• Tinker model pricing - Thinking Machines Lab
• Tinker platform - Thinking Machines Lab