Inkling AI Model 2026: Thinking Machines’ 975B Open-Weights

Thinking Machines Lab, the secretive startup founded by former OpenAI CTO Mira Murati, did something almost nobody expected from a lab this well-funded and this quiet: on July 15–16, 2026, it released the Inkling AI model — its first product and its first-ever open-weights release — straight onto Hugging Face for anyone to download. Inkling is a 975-billion-parameter mixture-of-experts with a 1M-token context window, one of the largest frontier-class models you can legally pull down and run yourself. That combination — genuine frontier scale, permissive weights, and a name-brand founder — makes this the most consequential open-weights drop since DeepSeek and Llama reset expectations. If you build with LLMs, it changes what “open” is allowed to mean in 2026.

What’s actually new with the Inkling AI model

Every major lab talks about openness. Thinking Machines Lab shipped it. Inkling arrived as downloadable weights under a permissive license, complete with a technical report, tokenizer, and reference inference code — not an API-only preview, not a “coming soon” waitlist. For a company that had said almost nothing publicly since its multi-billion-dollar raise, going from silence to a 975B open-weights model as the debut product is a deliberate statement of intent.

The architecture is a sparse mixture-of-experts. The headline 975 billion is the total parameter count across all experts; only a fraction — Thinking Machines reports roughly 50B active parameters per token — fires on any given forward pass. That trick makes a model this large tractable: you get the knowledge capacity of a near-trillion-parameter network at an inference cost closer to a mid-size dense model. Pair that with a 1M-token context window and you have a system that holds entire codebases, legal discovery sets, or book-length documents in working memory without retrieval gymnastics.

The other genuinely new thing is who is behind it. Murati’s team frames this open-weights release as a bet that frontier capability plus open access beats closed-lab gatekeeping — a direct shot across the bow of OpenAI, Anthropic, and Google. Whether Inkling matches GPT-5.6 on every benchmark matters less than the fact that a serious, well-capitalized lab now commits to putting frontier weights in your hands for free.

Why it matters

• Frontier scale is now downloadable. A 975B mixture-of-experts you can host yourself erodes the moat that closed labs built on “you can only rent this.” Sovereignty over your own model stack just got a lot more realistic.
• The 1M-token context window kills a lot of RAG. Many pipelines exist only to chunk and retrieve because context was scarce. When you can stuff a million tokens in directly, whole categories of infrastructure become optional.
• Sparse activation makes it affordable. Roughly 50B active parameters per token pushes serving costs and latency far below what “975B” implies. Small teams can realistically run this on rented multi-GPU nodes.
• Data privacy stops being a tradeoff. Regulated industries — health, finance, legal — can keep prompts and documents entirely on-premises instead of shipping them to a third-party API.
• It pressures pricing everywhere. A free, capable frontier model forces closed providers to justify their per-token rates. Expect the whole market to feel it.
• Mira Murati’s bet reshapes the narrative. A high-profile founder backing open weights gives the movement credibility with enterprises that were nervous about “community” models.

How to use the Inkling AI model today

You have two realistic paths: download and self-host if you have the hardware, or hit a hosted endpoint while you evaluate. Here’s the fast route.

1. Grab the weights. Pull the repo from Hugging Face. Expect a large download — plan storage accordingly.

   pip install -U huggingface_hub
   
   huggingface-cli download thinking-machines/inkling-975b-moe \
     --local-dir ./inkling --local-dir-use-symlinks False

2. Serve it with a modern MoE-aware runtime. vLLM handles sparse mixture-of-experts routing and long context better than a naive loader. Tensor-parallelize across your GPUs.

   pip install "vllm>=0.6.0"
   
   python -m vllm.entrypoints.openai.api_server \
     --model ./inkling \
     --tensor-parallel-size 8 \
     --max-model-len 1000000 \
     --gpu-memory-utilization 0.92 \
     --port 8000

3. Call it with the OpenAI-compatible API. vLLM exposes a drop-in endpoint, so existing SDK code works with only a base-URL change.

   curl http://localhost:8000/v1/chat/completions \
     -H "Content-Type: application/json" \
     -d '{
       "model": "./inkling",
       "messages": [
         {"role": "system", "content": "You are a precise technical assistant."},
         {"role": "user", "content": "Summarize the attached RFC in 5 bullets."}
       ],
       "temperature": 0.3,
       "max_tokens": 1024
     }'

4. Exploit the 1M-token context window. Instead of building a retrieval layer, feed the whole corpus in one request. In Python:

   from openai import OpenAI
   
   client = OpenAI(base_url="http://localhost:8000/v1", api_key="not-needed")
   
   with open("entire_codebase_dump.txt") as f:
       corpus = f.read()  # can be hundreds of thousands of tokens
   
   resp = client.chat.completions.create(
       model="./inkling",
       messages=[
           {"role": "system", "content": "Answer only from the provided source."},
           {"role": "user", "content": f"{corpus}\n\nWhere is auth handled, and what would break if I removed the session middleware?"},
       ],
       temperature=0.2,
   )
   print(resp.choices[0].message.content)

5. No GPUs? Use a hosted endpoint first. Several inference providers stood up Inkling within days of release. Point the same client at their base URL, evaluate quality on your real prompts, and decide whether self-hosting is worth it before you commit hardware.

One practical warning: a 1M-token prompt is not free even on an efficient model. Attention cost and time-to-first-token climb with input length, so measure latency on representative payloads and cache aggressively for repeated long-context calls.

How it compares

Benchmarks will churn for weeks, so treat exact numbers as provisional. The structural comparison — how Inkling stacks up as an open, sparse, long-context model against the closed frontier — is what matters for planning.

Dimension	Inkling (Thinking Machines)	GPT-5.6 (OpenAI)	Typical open dense model
Weights	Open, downloadable	Closed, API-only	Open
Total parameters	975B (MoE)	Undisclosed	70B–405B
Active params / token	~50B (sparse)	Undisclosed	All (dense)
Context window	1M tokens	Large, provider-set	128K–256K common
Self-hosting	Yes	No	Yes
Data stays on-prem	Yes	No	Yes
Cost model	Your compute	Per-token API	Your compute

The honest read on Inkling vs GPT-5.6: closed frontier models may still edge ahead on the hardest reasoning and tool-use evals, and they win on convenience — no GPUs to manage. Inkling’s advantage is control. You own the weights, your data never leaves your walls, and you’re not exposed to a vendor’s pricing or deprecation schedule. For many teams, that tradeoff now tips toward open.

What’s next

The ecosystem will form around Inkling faster than the model itself evolves. Within the first week, expect quantized builds (4-bit and 8-bit) that shrink the hardware bar, GGUF conversions for llama.cpp, and LoRA fine-tunes targeting code, medicine, and non-English languages. The community’s speed at wrapping a strong base model is exactly what turned earlier open releases into defaults, and Inkling has the scale to trigger the same effect.

The bigger question is Thinking Machines Lab’s cadence. Is this a one-time splash to plant a flag, or the start of a regular open-weights release train? Murati’s public framing points to a sustained open strategy, but the tell will be an Inkling-2, a smaller distilled variant for edge and consumer hardware, and — critically — whether the lab publishes enough training detail for the release to be reproducible rather than merely downloadable. Watch also for a license fight: “open weights” and “open source” are not the same thing, and the exact terms will shape what enterprises are willing to build on.

Finally, watch the incumbents react. A credible, free, frontier-scale open-weights release pressures closed labs to cut prices, extend context windows, or open more of their own stack. However Thinking Machines’ business model shakes out, the second-order effect — a more competitive, more open market — is the part that benefits everyone building today.

Frequently Asked Questions

What is the Inkling AI model?

Inkling is the first model from Thinking Machines Lab, Mira Murati’s startup, released as open weights on July 15–16, 2026. It’s a 975-billion-parameter mixture-of-experts with a 1M-token context window, making it one of the largest freely downloadable frontier-class models available.

Is Inkling really free to use?

The weights are released openly, so you can download and run them without a licensing fee. “Free” means no API bill, not zero cost — you pay for the compute to host it. Review the exact license terms before commercial deployment, since open-weights licenses vary in what they permit.

What does 975B mixture-of-experts actually mean?

975B is the total parameter count across all of the model’s experts. Because it’s a mixture-of-experts, only a subset — around 50B parameters — activates for any given token. You get the knowledge capacity of a near-trillion-parameter model at inference costs closer to a much smaller one.

Can I run the Inkling AI model on my own hardware?

Yes, if you have sufficient multi-GPU capacity. A full-precision 975B model needs a serious server node, but community quantized builds lower that bar considerably. If you lack hardware, hosted endpoints from third-party inference providers let you evaluate it first.

How useful is the 1M-token context window in practice?

Very — it lets you feed entire codebases, long document sets, or book-length inputs directly into a prompt instead of building a retrieval pipeline. Just budget for higher latency and memory use on very long inputs, and benchmark with realistic payloads before relying on it in production.

How does Inkling compare to GPT-5.6?

On the toughest reasoning benchmarks, closed frontier models like GPT-5.6 may still lead, and they’re easier to use since there’s nothing to host. Inkling’s edge is control: open weights, on-premises data, no per-token pricing, and no vendor lock-in. Which wins depends on whether you value peak capability and convenience or ownership and privacy.