Thinking Machines Kept Its Promise. Inkling 2 Small's Apache Weights Shipped in a Week, "Small" Needs 500 GB of GPU Memory, and Artificial Analysis Now Clocks It Even Slower.
Thinking Machines Lab updated its launch post on November 12: Inkling 2 Small's weights are on Hugging Face under Apache 2.0, not gated, in NVFP4 and MXFP8 variants, integrated with Unsloth, SGLang, vLLM, and TokenSpeed, with post-training through Miles, SkyRL, and Snowflake Cortex and API serving at NVIDIA, Fireworks, Together, Baseten, Nebius, Modal, and Cloudflare. The model card puts the NVFP4 checkpoint at 500 GB or more of aggregate GPU memory and MXFP8 at 800 GB or more. Artificial Analysis now lists it as an open-weights model at 31 on Intelligence Index v5.0.1, fourth of 28 in its comparison group, at $1.18 per Index task, and its November 13 recheck reads 29.6 minutes per task against 18.5 at launch and 67.71 seconds to first answer token against 34.15. We read the post, the card, Hugging Face, and AA.
By T.W. Ghost
Thinking Machines Lab updated its Inkling 2 Small launch post today, November 12, 2026: the weights are on Hugging Face. The repo thinkingmachines/Inkling-2-Small carries an Apache 2.0 license and is not gated, with an MXFP8 variant alongside the native NVFP4 checkpoint. The post lists integrations with Unsloth, SGLang, vLLM, and TokenSpeed, post-training through Miles, SkyRL, and Snowflake Cortex, and API serving at NVIDIA, Fireworks, Together, Baseten, Nebius, Modal, and Cloudflare. A week ago we wrote that this open-weights crown was a promise. The promise was kept, on schedule.
The catch is the word "Small." This is a 594-billion-parameter mixture of experts, and the model card says the NVFP4 checkpoint needs 500 GB or more of aggregate GPU memory, two B300s or four H200s, while the MXFP8 variant needs 800 GB or more, four B300s or eight H200s, for a gain the card puts at one to two points. The model was trained natively in NVFP4. None of that is hidden; it is a multi-GPU Blackwell or Hopper node by the vendor's own numbers, and "download" means something different at that size. We read the updated post, the model card, and the Hugging Face repo. Here is what held up.
What Shipped
| Inkling 2 Small weights (November 12, 2026) | |
|---|---|
| Where | Hugging Face thinkingmachines/Inkling-2-Small; Apache 2.0; not gated; separate MXFP8 repo |
| Formats | NVFP4, the native training format; MXFP8, one to two points better per the card |
| Hardware, NVFP4 | 500 GB or more aggregate GPU memory; 2x B300 or 4x H200 (model card) |
| Hardware, MXFP8 | 800 GB or more; 4x B300 or 8x H200 (model card) |
| Inference integrations | Unsloth, SGLang, vLLM, TokenSpeed |
| Post-training | Miles, SkyRL, Snowflake Cortex |
| API serving | NVIDIA, Fireworks, Together, Baseten, Nebius, Modal, Cloudflare |
| Third-party price | $0.30 in / $1.00 out per 1M on OpenRouter, Together, Cloudflare (third party) |
A week is a week. "Next week" on November 5 became November 12. In a month when Kimi K3.1 Pro's weights landed a day late under a custom license with a revenue clause, an Apache license on the date promised is the unusual part.
Half a terabyte of GPU memory is the entry fee. Two B300s or four H200s for the smaller format. A single-GPU user is not the audience; a provider is, which is why seven of them are on the serving list.
One third-party benchmark, labeled. Plotly ran its data-analytics bench and reports 39 of 43, or 91%, for $1.81, and places the model off its own cost-accuracy frontier. That is Plotly's bench and Plotly's frontier.
One claim we are not printing yet. Plotly's read of the launch post is that later training mixes used trajectories sampled from earlier models, including Kimi K3. We have not confirmed that in the post itself, and it would complicate "highest-scoring open model outside China," so it stays here as unconfirmed.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Weights on Hugging Face Nov 12; integrations; post-training partners; API providers | Verified (vendor post) |
| Apache 2.0; not gated; MXFP8 variant repo | Verified (Hugging Face API) |
| 500 GB and 800 GB; 2x B300 or 4x H200; 4x B300 or 8x H200; native NVFP4; one to two points | Verified (model card) |
| $0.30 / $1.00 on OpenRouter, Together, Cloudflare | Third party |
| Plotly 39 of 43, 91%, $1.81, off its frontier | Third party (Plotly) |
| Kimi K3 trajectories in later training mixes | Unconfirmed (Plotly's read); not printed as fact |
Who Should Use It
Download it if you operate a Blackwell or Hopper node with at least 500 GB of GPU memory and want the strongest Apache-licensed weights from outside China on AA's framing, scored 31.
Skip it on a single GPU, which the card rules out, and skip it for latency-bound work, where AA's recheck below went the wrong way.
The frontier is elsewhere: the open-weights leader on AA's board is still Kimi K3.1 Pro at 46, under a custom license, and the closed top is Claude Fable 5.2 at 65. Our model comparison has the current standings side by side.
Update, November 13, 2026
- AA flipped the label. Artificial Analysis now lists Inkling 2 Small as an open-weights model: 31 on Intelligence Index v5.0.1, fourth of 28 in its comparison group, $0.30 and $1.00 per million tokens, $1.18 per Index task, 154 million output tokens.
- And it got slower on AA's clock. Time per Index task now reads 29.6 minutes against 18.5 at launch; time to first answer token 67.71 seconds against 34.15; output speed 84 tokens per second. More providers did not make it faster on AA's measure.
Sources
- Thinking Machines Lab: Inkling 2 Small (updated launch post)
- Thinking Machines Lab: Inkling 2 Small model card
- Hugging Face: thinkingmachines/Inkling-2-Small
- Hugging Face: thinkingmachines/Inkling-2-Small-MXFP8
- Artificial Analysis: Inkling 2 Small model page
- Plotly: Inkling 2 Small on the Plotly data analytics bench (third party)
*Not sure whether open weights you cannot fit on one machine count as open for you? Take the free 2-minute quiz and get matched. Then read the launch post this one closes out: Inkling 2 Small at launch.*