Inkling Open-Weights Model Release
Inkling Open-Weights Model Release
Inkling is a 975 B‑parameter Mixture‑of‑Experts transformer with 41 B active weights, 1 M‑token context, and full open‑weights release, making it the first large‑scale multimodal open model that can be fine‑tuned on Thinking Machines’ Tinker platform.
Why Inkling matters: a multimodal, efficient, and customizable foundation model
Inkling’s combination of multimodal reasoning (text, images, audio, video), controllable thinking effort, and seamless integration with Tinker’s fine‑tuning UI gives developers a practical open‑weights base for building domain‑specific agents, even though it is not the strongest model on raw benchmark scores.
Core specifications
- Architecture: MoE transformer (256 routed experts per layer, 6 active per token) with sliding‑window + global attention (5:1 ratio) and relative positional embeddings.
- Scale: 975 B total parameters, 41 B active during inference; 1 M‑token context window.
- Training data: 45 trillion tokens spanning text, images, audio, and video.
- Multimodal inputs: Images as 40×40‑pixel patches processed by a four‑layer hMLP; audio as discrete dMel spectrograms.
- Effort control: A runtime "effort" knob (0.2‑0.99) lets users trade token usage for performance, achieving comparable scores to larger models with ~⅓ the token budget.
- Release: Full weights on Hugging Face (standard and NVFP4 checkpoints) and immediate availability on the Tinker console (64 K / 256 K context options).
Benchmark performance at effort = 0.99
Inkling is a balanced generalist rather than a specialist. On a shared 0‑100 scale it ranks competitively across several families:
| Category | Inkling score | Best open‑weight comparator |
|---|---|---|
| Reasoning (HLE + tools) | 46 % | Nemotron 3 Ultra (54 %) |
| Agentic coding (SWEBench Verified) | 77.6 % | Claude Fable 5 (95 %) |
| Vision (MMMU Pro) | 73.5 % | Gemini 3.1 Pro (82 %) |
| Audio (VoiceBench) | 91.4 % | Gemini 3.1 Pro (94 %) |
| Safety (FORTRESS Adversarial) | 78.0 % | Nemotron 3 Ultra (77.6 %) |
| Safety (StrongREJECT) | 98.6 % | All top models ≥ 98 % |
| Forecasting (Brier Index, no search) | 61.1 ± 0.79 | Kimi K2.6 (61.7 ± 0.54) |
The effort sweep shows Inkling reaching Nemotron 3 Ultra’s Terminal Bench 2.1 score with roughly one‑third the token count, illustrating its efficiency advantage.
Multimodal capabilities
- Audio: Speech‑to‑text, spoken‑instruction following, and long‑form audio reasoning. Scores of 91.4 % on VoiceBench place Inkling among the strongest open‑weight audio models.
- Vision: Image description, visual question answering, and chart/diagram reasoning. It handles visual‑plus‑code tasks via a lightweight Python tool for zoom/crop operations.
- Text & Code: Agentic tool use, one‑shot web‑app generation, and long‑form refinement loops (e.g., a multiplayer snake game refined over 40 iterations).
These capabilities are built on an encoder‑free design, simplifying the integration of modalities into the same transformer stream.
Controllable thinking effort
Inkling’s effort parameter lets developers balance latency and cost against performance. A sweep from 0.2 to 0.99 shows a smooth performance curve across Terminal Bench 2.1, HLE, and IFBench. For example, matching Nemotron 3 Ultra on Terminal Bench requires only ~33 % of the tokens, directly reducing inference cost for high‑throughput applications.
Safety and epistemics
Inkling was trained with a safety spec covering CBRN, cyber, and manipulation threats, and it was evaluated by external testers. It achieves the highest FORTRESS adversarial refusal rate (78 %) among open‑weight models and exceeds 98 % on StrongREJECT, indicating strong refusal of harmful requests while preserving benign answer rates.
Calibration (epistemics) was improved via RL with rubric and factuality graders, plus abstention‑aware rewards that encourage “I don’t know” when confidence is low. On ForecastBench, Inkling’s Brier Index (61.1) rivals the best open models, suggesting reliable uncertainty estimation.
Inkling‑Small preview
A lighter sibling, Inkling‑Small, uses 276 B total parameters (12 B active). It matches or exceeds the larger model on many benchmarks (e.g., IFBench 83.4 % vs. 79.8 %) while offering lower latency and cost, making it suitable for workloads where speed is critical.
Customization via Tinker
Inkling can be fine‑tuned directly in the Tinker console, which now includes an "Inkling Playground" for interactive chatting and tool use. The release demonstrates self‑fine‑tuning: Inkling generated its own fine‑tuning job, executed it, and evaluated the result, showcasing the platform’s end‑to‑end workflow.
Developers can access the model through APIs on Together, Fireworks, Modal, Databricks, Baseten, and via open‑source runtimes such as vLLM, TokenSpeed, and llama.cpp (including an NVFP4‑optimized checkpoint for NVIDIA Blackwell GPUs).
Community reaction (Hacker News highlights)
- Positive reception: Commenters praised the multimodal breadth, especially audio support, and the fact that an American lab is delivering a competitive open model.
- Benchmark skepticism: Some users noted that Inkling lags behind GLM‑5.2 on raw scores but highlighted its unique combination of features (multimodality, effort control, Tinker integration) as differentiators.
- Practical concerns: Users asked about local deployment (e.g., llama.cpp ports) and licensing (Apache‑2.0 with an additional AUP).
- Future expectations: The community anticipates a model family (small + large) and sees Inkling as a potential cornerstone for enterprise‑grade fine‑tuning services.
Takeaways
- Inkling is the first large‑scale open‑weights model that natively handles text, images, audio, and video with a 1 M‑token context.
- Its controllable effort mechanism provides a clear cost‑performance trade‑off, outperforming larger competitors on token efficiency.
- Integrated fine‑tuning on Tinker makes Inkling immediately usable for domain‑specific applications, positioning it as a practical foundation model rather than a pure benchmark champion.
Where to get Inkling
- Weights: Hugging Face repository
thinkingmachines/inkling(standard and NVFP4 checkpoints). - Interactive demo: Inkling Playground in the Tinker console.
- API access: Available via Together, Fireworks, Modal, Databricks, Baseten, and open‑source runtimes (vLLM, TokenSpeed, llama.cpp, etc.).
Bottom line: Inkling’s multimodal breadth, efficient effort control, and out‑of‑the‑box fine‑tuning pipeline make it a compelling open‑weights foundation for developers who need a customizable, cost‑effective model, even if it is not the absolute top performer on every benchmark.