Inkling 2 Small Is Here: 594B Parameters, $0.30/$1 Pricing, Open Weights Still Pending

Thinking Machines Lab shipped Inkling 2 Small on November 5, 2026: a 594-billion-parameter mixture-of-experts model (17B active per token) that scores 31 on the Artificial Analysis Intelligence Index and runs at $0.30 per million input tokens and $1.00 per million output tokens on Tinker. The weights aren't out yet. The company says a full Apache 2.0 release is coming "soon," pending external safety testing.

That gap between "available today" and "open next week" is the whole story here. Mira Murati's lab has made a habit of promising open weights and then making you wait, and this release follows the same script.

What exactly did Thinking Machines release?

Inkling 2 Small is a 48-layer decoder-only transformer with sparse MoE feed-forward layers, hybrid local/global attention, and a vision transformer bolted on for image input. It takes text and images, outputs text, and runs with variable "effort" levels you can dial up for harder problems.

The headline specs:

• 594B total parameters, 17B active per forward pass
• Context window: 64K tokens standard, 256K via the :peft:262144 suffix on Tinker (the model card claims the architecture supports up to 1M tokens)
• Pretraining: four phases, 12.2 trillion tokens, trained entirely in NVFP4 precision
• Post-training: a unified reinforcement learning run spanning 9 days across 2,048 GPUs, 17 million rollouts
• License today: Tinker Terms of Use (closed weights). Apache 2.0 planned after safety testing wraps

It's live right now for inference and fine-tuning through Tinker and the Tinker Playground. You can also let it drive its own fine-tuning workflow, which is a neat trick but not something I'd hand production budget to without watching it closely first.

How does it actually perform?

The jump from the previous generation is real. Inkling 2 Small's AA Intelligence Index score of 31 is more than double its predecessor's 13, and the gains are lopsided toward agentic work:

That AA-Briefcase score of 1,527 beats Claude Sonnet 5.5's 1,439, which is a genuinely good number for a model this cheap. Among models developed outside China, Thinking Machines is calling this the strongest open-weight release once the weights land, ahead of Mistral Large 4 (26 on the same index) and Meta's Muse Spark Open (25). It also clears Reflection AI's Beam and Nvidia's Nemotron by a wide margin.

It's not winning everywhere, though. DeepSeek-V4.1-Flash still edges it out on FrontierCode (28.5% vs 25.4%), and on overall intelligence Inkling 2 Small lands roughly where Gemini 3.8 Flash sits. If you're picking a model purely on leaderboard position, this is a strong second-place finish against the field, not a knockout.

What does it cost?

Pricing runs through Tinker and it's genuinely cheap for what you get:

There's a catch worth flagging before you budget around these numbers. Per-task cost jumped to $1.18, up from roughly $0.20 for the predecessor, because the model now generates about 150,000 output tokens per task on average, four times as many as before. The AA-Briefcase comparison with Claude Sonnet 5.5 used "far cheaper, $7.77 per task for Gemini 3.8 Flash vs this" as the selling point, and that's true on paper. But if your workload is token-hungry agentic chains, that per-task figure matters more than the per-million-token sticker price. Do the math on your actual task length before assuming this undercuts everything else on the market.

One genuinely good number: hallucination rates dropped, partly because the model now attempts fewer questions (53% vs 77% previously). That's the model getting better at saying "I don't know" rather than guessing, which I'd take any day over a confident wrong answer.

Is Inkling 2 Small actually small?

No, not really, and Hacker News called this out immediately. At 594B total parameters, this thing is bigger than the original Inkling Small from July 2026 (276B params) and sits well above what most people would call deployable on a single node. One commenter put it bluntly: "This isn't small at all, for most budgets." The "small" in the name is relative to Thinking Machines' flagship Inkling (975B), not relative to what you can actually run on your own hardware.

There's a second thread of skepticism worth taking seriously: some commenters pointed out the training data likely leans on earlier Kimi and other proprietary model outputs, meaning a chunk of the intelligence here is distilled from compute Thinking Machines didn't pay for directly. That's common practice across the industry at this point, but it's a fair asterisk on any "we built the best Western open model" claim.

I'd summarize the HN reaction as: impressive engineering velocity, genuinely useful benchmark gains, but don't take the marketing framing (smallest, cheapest, most open) at face value without checking the fine print on each claim separately.

When do the weights actually show up?

"Next week," according to the launch post, pending results from external safety testing. Thinking Machines has done this sequence before: announce on Tinker first, open the weights a bit later once red-teaming clears. If you need Apache 2.0 weights today for self-hosting, you're out of luck. If you're fine using it through Tinker's API in the meantime, there's no wait at all, and the fine-tuning access is arguably the more interesting part of the release anyway since that's the actual product Tinker is built around.

Should you use it now?

If you're already on Tinker and want a vision-capable model with strong agentic and office-task benchmarks at a low per-token price, there's no reason to wait. The coding and automation numbers are good enough to justify a real evaluation against whatever you're running today, especially DeepSeek-V4.1-Flash or Gemini 3.8 Flash if those are your current baseline.

If you were specifically waiting for self-hostable, Apache-licensed weights to run this on your own GPUs, you're still waiting, and "next week" from an AI lab is a number I've learned not to treat as a hard date.

Watch the announcements

For more on the models Inkling 2 Small is actually competing against, see our breakdowns of DeepSeek V4.1 Flash, Gemini 3.8 Flash, Meta's Muse Spark Open, and Kimi K3.1 Pro.