Inkling
Thinking Machines Lab‘s first open-weights model, released 2026-07-15 under Apache-2.0 — a 975B-parameter Mixture-of-Experts with 41B active, pretrained on 45 trillion tokens of text, images, audio and video, with a 1M-token context inkling-announcement. A smaller sibling, Inkling-Small (276B total / 12B active), is still in testing and aimed at cost- and latency-constrained work.
The lab is unusually blunt about where it sits: “Inkling is not the strongest overall model available today, open or closed.” The pitch is instead a base to customize — multimodal input, cheap thinking, and fine-tuning on the lab’s own Tinker platform inkling-announcement.
Architecture
A 66-layer decoder-only transformer over a sparse MoE backbone inkling-announcement:
- 256 routed experts + 2 shared experts per MoE layer, 6 active per token — the shared-expert variant deepseek popularized (moe-architecture).
- Sigmoid router with auxiliary-loss-free load balancing.
- Interleaved sliding-window and global attention, 5:1.
- Relative positional embeddings, not RoPE — a deviation from the field’s default.
- Encoder-free vision and audio: 40×40 pixel patches through an hMLP encoder; dMel spectrograms for audio, projected into the same hidden space as text.
Post-training leaned on synthetic-data bootstrapping and 30M+ RL rollouts on reasoning tasks.
Where it ranks
Self-reported at effort=0.99 inkling-announcement — no independent artificial-analysis re-run
yet, so read these as vendor numbers (llm-benchmarks):
| Axis | Benchmark | Score |
|---|---|---|
| Reasoning | AIME 2026 | 97.1% |
| Reasoning | GPQA Diamond | 87.2% |
| Reasoning | HLE (with tools) | 46.0% |
| Coding | SWE-bench Verified | 77.6% |
| Coding | Terminal Bench 2.1 | 63.8% |
| Vision | MMMU Pro | 73.5% |
| Audio | VoiceBench | 91.4% |
| Factuality | SimpleQA Verified | 43.9% |
| Safety | FORTRESS (adversarial) | 78.0% |
Competitive among open weights, behind the closed frontier (GPT-5.6 Sol, Claude Fable 5) on several. The 43.9% on SimpleQA Verified is the weak spot — this is not a model to ask facts of.
Running it
Weights are on Hugging Face in BF16, MXFP8 and NVFP4. BF16 wants 2TB+ of aggregated VRAM (8× B300 or 16× H200); the NVFP4 checkpoint drops that to ~600GB inkling-announcement. So “open weights” here means a GPU cluster, not a laptop — the opposite end of the footprint axis from gemma-4 at 16GB (quantization). Served by TogetherAI, Fireworks, Modal, Databricks and Baseten; runs on SGLang, vLLM, TokenSpeed, and llama.cpp via Unsloth. Fine-tuning is on Tinker, at a limited-time 50% discount (rates not published on the announcement page).
Why it matters here
Two things the spoke has been tracking (synthesis).
A US open-weight lab that isn’t a hyperscaler. Willison‘s read inkling-willison-review: “It’s good to see the US open weights ecosystem gain a new viable contender to join NVIDIA Nemotron and Gemma 4.” The open field has been Chinese-led; a US startup shipping Apache-2.0 at this scale is a different shape from google‘s dual-track gemma-4 hedge.
Customization as the product, not rank. Every other model here competes on the capability × price × context board. Inkling concedes the board and sells the base you fine-tune — with the lab’s revenue sitting in Tinker, one layer up. The weights are the funnel.
Caveats
- Benchmarks are self-reported and dated 2026-07-16; treat as a snapshot.
- Training data is barely documented — “public internet content,” “publicly accessible data repositories,” third-party datasets, with no breakdown inkling-willison-review. Apache-2.0 on the weights says nothing about what went in.
- The model card admits residual safety risk: an “occasional tendency to comply with role-play and indirectly framed prompts concerning harmful topics,” with Llama Guard-style moderation recommended inkling-announcement.
Related
thinking-machines-lab · tinker · inkling-announcement · inkling-willison-review · open-weight-models · moe-architecture · gemma-4 · llm-benchmarks · quantization · simon-willison