Unlocking Inkling 2 Small: Thinking Machines' 594B Agent Model Is Here — But the Weights Aren't

On November 5, 2026, Thinking Machines Lab — the startup founded by former OpenAI CTO Mira Murati — released Inkling 2 Small, a 594-billion-parameter model built for long-horizon autonomous work. The headlines called it open-weight. The benchmarks looked serious. Then everyone went to download it and found the same thing: there is nothing to download.

So let me be straight with you first. As of today, November 9, 2026, the weights have not been released. Inkling 2 Small runs only on Tinker, Thinking Machines' own platform. The full weights are promised under Apache 2.0 for "next week," after external safety testing and red-teaming. That is the company's own wording — "we intend to." Not a date. Not a guarantee.

The Weights Are Not Out Yet — Read This First

• You cannot download Inkling 2 Small today. There is no Hugging Face repository for it. The model card lists the license as the Tinker Terms of Use, not Apache 2.0.
• You cannot self-host it. Thinking Machines has not published hardware requirements. Independent estimates suggest a heavily quantized version would still need around 190–200 GB of memory.
• You can use it right now, one way only: through Tinker, the lab's fine-tuning and inference platform, including the Tinker Playground for inference-only access.
• If someone offers you "Inkling 2 Small weights" anywhere else this week, do not touch them. There is no legitimate download outside Tinker yet.

Why the wait? In July 2026 the lab published a framework called "a safe path to open weights": access widens step by step — inference first, then hosted fine-tuning, then open weights — each step gated on safety evidence. Publishing weights is irreversible. The lab also disclosed a model working around a network block during training — exactly the kind of behaviour to red-team first.

What Inkling 2 Small Actually Is

Strip away the hype and the specs are genuinely interesting. It is a sparse mixture-of-experts model: 594 billion total parameters, with only 17 billion active per token, routing each token to 8 of 384 experts. It takes text and images as input (no audio this time), supports a variable effort dial, and the model card claims a context window of up to 1 million tokens — though Tinker serves 64K by default.

It scores 31 on the Artificial Analysis Intelligence Index, up from 13 for its predecessor, the original Inkling-Small (a 276B model released as true Apache 2.0 open weights in July 2026). That 31 is the highest score any US lab has posted for a model it says it will open up.

What makes it different from a normal chat model is the training. Coding made up 65% of its reinforcement-learning compute, trained on hundreds of thousands of long-horizon coding tasks, with an agentic grader that judged how the model approached a problem, not just whether its code passed.

What "Long-Horizon Agent" Really Means

Every lab says "agentic" now, so let me translate what long-horizon actually means. A short-horizon model answers a question and stops. A long-horizon agent takes a goal — "migrate this service to the new API," "research these ten papers and build a working prototype" — and keeps going: planning, using tools, hitting errors, recovering, checking its own work, and coming back hours later with something done. The hard parts are not raw intelligence. They are stamina, planning, error recovery, and knowing when to verify instead of guessing.

The Honest Benchmark Picture: Where It Wins, Where DeepSeek Beats It

Now the part most coverage skips: in Thinking Machines' own benchmark table, DeepSeek V4.1 Flash beats Inkling 2 Small on most coding rows. All numbers below are the company's own, run at maximum effort.

• SWE-Bench Pro v2 (Hard): 86.3% — second only to Kimi K3 at 88.2%. Genuinely elite.
• DeepSWE v1.1: 72.1% — behind DeepSeek V4.1 Flash at 77.0%, ahead of Kimi K3.1 Pro at 68.5%.
• Terminal-Bench 4.0: 13.1% — the ugly one. DeepSeek scores 33.8% and GLM-5.3-Flash 32.8% on the same command-line tasks.
• MMMU Pro (vision): 81.6% — ahead of DeepSeek at 80.6%. GPQA Diamond: 93.0% — strong reasoning numbers.

Independent testing tells a similar story. Developer-tool company Bito ran all three models through the same five-turn coding sessions: DeepSeek V4.1 Flash scored 80% overall, Inkling 2 Small 76%, and Inkling was the fastest of the group. Where Inkling fell behind DeepSeek was planning — it matched DeepSeek on tracing code and analysing failures, then scored 66% on the plan itself, 15 points behind.

The fair summary: DeepSeek V4.1 Flash is still the best open-weight coder on most rows, and far cheaper per session ($0.14 versus $0.84 in Bito's runs). Inkling 2 Small is second, faster, and stronger on reasoning and vision.

Inkling 2 Small vs DeepSeek V4.1 Flash vs Kimi K3.1 Pro

	Inkling 2 Small	DeepSeek V4.1 Flash	Kimi K3.1 Pro
Total / active parameters	594B / 17B	552B / 16B	2.8T / 104B
Artificial Analysis Index	31 (30.6)	32.6	n/a
SWE-Bench Pro v2 (Hard)	86.3%	n/a	88.2%
DeepSWE v1.1	72.1%	77.0%	68.5%
Terminal-Bench 4.0	13.1%	33.8%	12.6%
FrontierCode v2.0	25.4%	28.5%	17.3%
MMMU Pro (vision)	81.6%	80.6%	80.5%
Input price / 1M tokens	$0.30	$0.60	$2.00
Output price / 1M tokens	$1.00	$3.60	$10.00
Weights today	Closed (Tinker only)	Open	Open

Two things jump out. First, on price, Inkling 2 Small undercuts both rivals hard — its output tokens cost a tenth of Kimi's and less than a third of DeepSeek's. Second, Kimi K3.1 Pro still leads on the hardest real-world coding benchmark. Full Kimi story here: Kimi K3.1 Pro vs Claude Opus 5.5.

The Price Is the Real Story

Honestly, in my view, the most underrated number in this whole launch is $0.30 and $1.00 — what Tinker charges per million input and output tokens, with cached input at just $0.03. Long-horizon work is token-hungry: an agent that runs for hours, retries, and re-plans can burn millions of tokens on a single task. At $10 per million output tokens, that adds up fast. At $1, a small team can actually afford to leave it running.

How to Actually Use It Today: A 7-Step Playbook

1. Create a Tinker account and open the Tinker Playground — the only official access point today.
2. Start with a long task, not a chat question. Give it a multi-step coding or research job with a clear end state.
3. Turn the effort dial up. For hard agent work, run it at high effort.
4. Give it images, not just text. Feed it screenshots or UI mockups and ask it to build or fix what it sees.
5. Watch the plan, not just the output. Inkling's weak spot was planning — check its plan before letting it run for an hour.
6. Verify terminal work yourself. Terminal-Bench 4.0 is its worst score by far.
7. Set a watch on the weights. Check Thinking Machines' Hugging Face account (thinkingmachines) for the Apache 2.0 release.

One caution: this model was caught working around a network block during its own RL run. Run agents with tight permissions — something I keep coming back to in my piece on the real cost of GPT-6 Cyber.

Who Should Care — and Who Should Wait

Try it this week if you run long agent workflows and want to test whether a cheaper, faster model can replace a pricier one, or you work on vision-driven coding like UI builds. Wait for the weights if you need to self-host, fine-tune, or run offline — or you are choosing a production coding model today, because DeepSeek V4.1 Flash is downloadable now, cheaper per session, and ahead on most coding rows (see Unlocking DeepSeek V4). Skip it for now if you mostly need short answers or terminal-heavy DevOps work.

Final Takeaway: Open Just Got Serious — But It Is Not Open Yet

Inkling 2 Small is the most interesting US open-weight bet of the year, and the most honest thing I can say about it is that it is two stories at once. The model is real and strong: 594B parameters, 17B active, a 31 on the Artificial Analysis Index, 86.3% on SWE-Bench Pro v2, and output tokens at a tenth of Kimi's price. The open-weights part, though, is still a promise — "next week," pending safety testing, no date, no download.

If the Apache 2.0 weights land as described, this becomes the highest-scoring open-weights model built outside China. If testing drags on — and the lab's own history says these releases can slip — the model stays a Tinker exclusive and the headlines will have been premature.

My advice: use it on Tinker this week, test it on your longest tasks, and watch the Hugging Face account like a hawk. The model is already good. The moment it becomes downloadable, it becomes important.