Inkling: Murati’s Lab Bets on Customizable Open Bases, Not SOTA Chat
Thinking Machines Lab released Inkling—a 975B MoE open-weight multimodal (41B active, 1M context)—on Hugging Face with fine-tuning via Tinker. AA ~41 as leading U.S. open weights at debut; lab admits it is not the strongest model available.
A Murati-led lab choosing open weights first—and admitting it is not the strongest model available—reframes the U.S. open-source story from “Meta Llama only” to a second pillar: customizable multimodal bases with a commercial fine-tune stack. That is a structural competitor to both rent-only frontier APIs and pure weight dumps without ops tooling.
Thinking Machines Lab on July 15, 2026 released Inkling—its first foundation model—as open weights. The company primary describes a Mixture-of-Experts transformer with 975B total parameters, 41B active, a context window up to 1M tokens, and pretraining on 45 trillion tokens of text, images, audio, and video. Weights are on Hugging Face; fine-tuning ships on Tinker, with API/inference partners including Together, Fireworks, Modal, Databricks, and Baseten.
What Inkling is
From the July 15 Thinking Machines post (“Inkling: Our open-weights model”) and corroborating TechCrunch, Axios, WSJ, and Fortune coverage the same day:
| Spec | Detail |
|---|---|
| Architecture | MoE transformer; ~DeepSeek-V3-style expert routing (256 routed + 2 shared; 6 active routed) |
| Scale | 975B total / 41B active parameters |
| Context | Up to 1M tokens (Tinker options at 64K / 256K at launch) |
| Modalities | Native text, images, audio; video in pretrain mix |
| Training | 45T tokens; hybrid Muon/Adam; large-scale async RL (>30M rollouts cited) |
| Companion | Inkling-Small preview (276B total / 12B active)—weights later after testing |
| Positioning | Broad generalist for fine-tuning; controllable thinking effort |
The company demoed an agentic loop in which Inkling fine-tunes itself on Tinker (lipogram experiment: no letter “e”), underscoring the product thesis: customization over raw leaderboard rank.
Claims vs checks
Thinking Machines does not sell Inkling as a closed chat subscription rival to ChatGPT or Claude. Revenue and GTM center on Tinker—enterprises (including Bridgewater, per contemporaneous Fortune reporting) that want to post-train on private data. Valuation context from press: roughly $12B-class after earlier fundraising.
| Source | Snapshot | Standing |
|---|---|---|
| Artificial Analysis Intelligence Index | AA post Jul 15, 2026 | 41 — leading U.S. open-weights model on AA at debut (above Nemotron 3 Ultra at 38, Gemma 4 31B at 29, gpt-oss-120b at 24). Not a closed-frontier rival: well below Fable/Sol/Kimi K3 (50+ club) |
| AA agentic slices | Same | GDPval-AA v2 Elo 1238 (above Kimi K2.6 1190 and DeepSeek v4 Flash max 1189); 24% on 𝜏³-Banking; strong token efficiency (~25K output tokens per Intelligence Index task vs ~37–43K for several open peers) |
| Arena Text Overall | Board dated Jul 16, 2026 | inkling (listed under Thinky) #67 at 1441 ±15 (~1,543 votes) — early, wide confidence band; open-weight multiparty mid-table, not top-20 chat preference |
Inkling’s independent story is best U.S. open base for customization, not Arena #1 chat—matching the lab’s own framing.
Limits
- AA 41 and Arena #67 are launch-window snapshots; vote bands and Index updates move.
- Lab training/rollout figures (45T tokens, >30M rollouts) are vendor-claimed.
- Inkling-Small weights were not public at debut.
- ~$12B valuation is press context, not a primary lab disclosure in the model post.
Sources
- Thinking Machines Lab: “Inkling: Our open-weights model” (July 15, 2026) — primary specs, architecture, Tinker availability, Hugging Face.
- Artificial Analysis: “Thinking Machines has released Inkling, the new leading U.S. open weights model” (July 15, 2026).
- Arena.ai Text Overall leaderboard snapshot (July 16, 2026).
- TechCrunch, Axios, Wall Street Journal, Fortune / Bloomberg same-day coverage of Inkling’s open-weight launch and enterprise positioning (July 15, 2026).