AI News · 4 min read
Inkling 2 Small open weights land on Hugging Face
On November 12, 2026, Thinking Machines published Inkling 2 Small on Hugging Face under Apache 2.0. The native checkpoint needs at least 500GB of VRAM.
What Thinking Machines shipped on November 12
On November 12, 2026, Thinking Machines updated Introducing Inkling 2 Small ↗: the weights are on Hugging Face. The same note says the model is integrated with Unsloth, SGLang, vLLM, and TokenSpeed, can be post-trained with Miles, SkyRL, and Snowflake Cortex Training, and is available via API at NVIDIA, Fireworks, Together, Baseten, Nebius, Modal, and Cloudflare.
The picture above is ours.
The November 5 launch, when the weights were still "next week," is our earlier post. Two checkpoints: Inkling-2-Small ↗ (NVFP4) and Inkling-2-Small-MXFP8 ↗. The model card ↗ lists the license as Apache 2.0.
Who can actually run the files
The model card is blunt about the machine. This is a 594B total / 17B active mixture-of-experts: a 48-layer decoder, 8 of 384 experts per token, hybrid local and global attention, image inputs, text out, and a context window of up to 1 million tokens.
| Checkpoint | Aggregated VRAM | Example the card gives |
|---|---|---|
| NVFP4 (native) | At least 500GB | 2x B300 or 4x H200 |
| MXFP8 | At least 800GB | 4x B300 or 8x H200 |
Thinking Machines trained it in NVFP4 and tells you to use that file. MXFP8 gains 1–2 percentage points on evals, which the lab says does not matter in practice. vLLM's November 12 post ↗ says the mixed-precision experts can sit on one node. That node is still the cluster in the table.
The cheap path is an API
Plotly ↗ (November 12) says OpenRouter, Together, and Cloudflare serve it at $0.30 per million input tokens and $1.00 per million output tokens, with a 1 million-token window. Ollama's cloud tag ↗ lists the same $0.30 input and $1.00 output, plus $0.03 cached input. The tag is inkling-2-small:cloud (ollama run inkling-2-small:cloud). Ollama's tags page shows that cloud tag, not a local weight.
Plotly also writes that a 4-bit copy needs about 356GB and fits a 512GB Mac Studio. The model card does not. Use the 500GB line.
vLLM's agent test, tensor parallel 8 with DSpark and prefix caching, reports about 1,070 output tokens per second per GPU at a 32K session and about 450 at 256K, at a concurrency where each user still gets 100 tokens per second. Those figures are vLLM's. Plotly's reading of the lab's report says later training mixes included trajectories from earlier models, including Kimi K3. On Plotly's private exam, one run scored 39 of 43 for $1.81. Plotly says other models moved by up to three points between runs.
Where it sits against weights we already covered
Every score in this table is from Thinking Machines' model card, not our lab. The announcement prose still cites an Artificial Analysis Intelligence Index of 31. The card's comparison table lists AA Index v5.0 at 30.6.
| Model, on their table | AA Index v5.0 | SWE-bench Pro v2 (Hard) |
|---|---|---|
| Inkling 2 Small | 30.6 | 86.3% |
| DeepSeek-V4.1-Flash-1017 | 32.6 | 89.7% |
That DeepSeek row is not the speed tier. DeepSeek V4.1 Flash Ultrafast on Vercel's AI Gateway is $2.40 in and $9.60 out per million tokens. ERNIE-5.5-Flash-Base is 280B total and 14B active, under Baidu's custom license, and the file is the end of pre-training. Qwen4-Max preview has no weights, is Token Plan only, and ends November 25. The map is best open-source LLMs 2026.
What people are saying
"MoE 594B Active 17B (definitely not that small)" — Aritra (@ariG23498) ↗, at Hugging Face
"the "small" part: 594B total params, 17B active, and the native NVFP4 checkpoint wants at least 500GB of VRAM. so 2x B300 or 4x H200 😅" — Shaoze Chen ↗
"Inkling 2 Small from @thinkymachines is live on AI Gateway as an open weights model." — Vercel Developers ↗
"One week from announcement to open weights. That's new." — Adrian Blake ↗
"About 450 tokens a second per GPU. That's where Inkling 2 Small landed at 256K context in vLLM's own agent test, down from roughly 1,070 at 32K." — Deepu ↗, describing the vLLM test above
Reddit was not reachable for this story.
Our take
Call an API this week. Do not buy a laptop for the weights. The model card's floor is 500GB of aggregated VRAM for NVFP4, or 800GB for MXFP8. Shaoze Chen's list matches the card. Aritra's joke matches the name: "Small" is 17B active, not the machine. Download only if you already have that cluster and you need the Apache 2.0 file for post-training with Miles, SkyRL, or Snowflake Cortex. Everyone else should use the $0.30 / $1.00 hosts, or Ollama's cloud tag.
Do not read the card as a win over DeepSeek. On Thinking Machines' table, DeepSeek-V4.1-Flash-1017 leads on the AA index and on SWE-bench Pro v2 Hard. Inkling's case is the license, day-0 serving, and a lower token price than the Ultrafast tier. Test your own harness. The coding guide and the coding-assistant benchmark are for that, not a vendor row. Qwen4-Max is not a file you can keep. Plotly's 39 of 43 is one private exam. It is a reason to spend $1.81, not a reason to stand up eight H200s.