Inkling 2 Small: Thinking Machines kept its one-week promise to the day. The weights are out, and its own page still says "pending"

Updated on 14/11/2026: Thinking Machines has published the promised post on its 4-bit training recipe; Artificial Analysis now ranks Inkling 2 Small 4th of 29.

Thursday 05/11/2026, Thinking Machines introduces Inkling 2 Small and sets a date: the weights, the billions of numbers that are the model, will be released "next week", once external safety testing is done. Thursday 12/11/2026, at 19:53 UTC, the files land on Hugging Face. One week, to the day. The question is what you can do with them: 393 gigabytes, a score of 31 measured by an independent referee, and a license that lets you copy, modify and redistribute it without paying a cent.

In short

• The model: a mixture of experts with 594 billion parameters, 17 billion of which work on each token (a piece of a word), reading text and images; weights under the Apache 2.0 license since 12/11/2026.
• The score: 31 on the Artificial Analysis index, 4th of 29 open-weights models above 150 billion parameters on 14/11/2026; the first Inkling scored 12.
• The price without a machine: $0.30 per million tokens sent, $1.00 per million tokens produced, at the four hosts whose page I could read.

Seven days behind the glass

I had met the first Inkling as a teacher: Milky told how Nous Research used its answers to re-educate Qwen in Arete-27B. The teacher moves to version 2, and this time the lab changed its method.

On 05/11/2026, Inkling 2 Small only exists on Tinker, the in-house platform where you can query it and retrain it without ever touching the files. That is the recipe Thinking Machines published on 31/07/2026 under the title "A Safe Path to Open Weights": on a platform, "the provider can monitor usage, maintain guardrails, and revoke access". Once the weights are public, none of that is left. The same text says it in one word: "release is irreversible".

I opened the repository history on Hugging Face. The repository has existed since 04/11/2026, but the first file upload in its history is dated 12/11/2026 at 19:53:15 UTC, the heavy version one second later. Promise kept on the date.

On what the tests found, it gets thinner. The card mentions "open-ended external red-teaming" and concludes the model "did not present risk of material uplift beyond what's already available in the open-weight ecosystem". The post shows scores credited to Scale AI, Apollo Research and Goodhart Labs. A separate report, with who tested what? Nothing in the lab's News and Blog sections, checked on 13/11/2026 between 07:52 and 07:55 UTC. And the detail that made me smile: in the same post as the 12/11/2026 box announcing the weights, the safety section still says "The release of Inkling 2 Small’s weights is pending further safety evaluations by us and our partners".

Why test at all? The post tells it itself: early checkpoints, facing hard tasks, were, in some cases, "fabricating results, ignoring instructions, and concealing what they had done"; later, failing at a task, the model in training "worked around a network block through a third-party service". The lab paused the run.

What came out of the vault

Two repositories. The main one, in NVFP4, the 4-bit format the model was trained in: 393.5 gigabytes of weight files. The MXFP8 variant, in 8 bits: 622.8 gigabytes (our calculation, from the sizes reported by Hugging Face). On 13/11/2026, the lab published on its blog the post promised on this recipe, by Jack Cook with others at the lab. According to the post, Inkling 2 Small is the largest model the lab is aware of that was trained end to end with a native 4-bit forward pass; the numerics are configured the same way from the start of pretraining through the final reinforcement-learning rollout. The 4 bits are used for the forward pass of the expert layers; the backward pass, the one that updates the weights, runs in MXFP8, on 8 bits. In the lab's largest jobs, the recipe gives an 8% wall-clock speedup in pretraining and a 20% speedup in reinforcement learning over full MXFP8 training. In pretraining Inkling 2 Small it costs about 2% in compute efficiency: the model reaches the loss that MXFP8 training would have reached with about 2% less compute. The lab says it trains on NVIDIA's Grace Blackwell Ultra platform, the GB300.

Storing them is not enough. To run it at home, the card is blunt: at least 500 gigabytes of aggregated graphics card memory for the 4-bit version, for example two NVIDIA B300 cards or four H200s; 800 gigabytes for the 8-bit version. Data-centre cards, not living-room ones. The lab recommends the 4-bit version "for all use-cases": the other gains "1–2 percentage points" on its tests, a gap it does not think is "meaningful for any practical use-cases".

The score, checked with the referee

The post announced 31 on the Artificial Analysis index. I went to look. This index combines nine tests the site runs on the models itself, from a business file to finish to programming tasks at the command line; higher is better. On 13/11/2026 at 07:52 UTC, the page does show 31 (30.6 before rounding), 4th of 28 open-weights models above 150 billion parameters. On 14/11/2026 at 01:37 UTC, the page still shows 31 and ranks Inkling 2 Small 4th of 29 models in that class.

For scale, at the same referee: the top score shown is 65, for Claude Fable 5.2 at its maximum setting. Kimi K3.1 Pro, another open model, is at 46; Gemini 3.8 Flash, closed, at 31 on its high setting. And the first Inkling, at 12. The small one of the second generation takes the score from 12, the first Inkling's, to 31. The price is time: 29.6 minutes on average per index task on the lab's own API, against 14.6 for the median of comparable models. On 14/11/2026 at 01:37 UTC, the page gives 28.1 minutes against 14.3 for the median. The post's other scores are published by the lab, which says it uses "externally reported evaluations" when available.

The price, if you don't have 500 gigabytes at hand

Together, Modal, Cloudflare and Tinker show the same rate: $0.30 per million tokens sent to the model, $1.00 per million tokens it writes. In practice, you give it a 100,000-token file and it hands back a 10,000-token summary: $0.03 for the reading, $0.01 for the writing, $0.04 in all (our calculation). On Tinker, the first Inkling costs $1.00 per million tokens sent and $4.05 per million tokens written for serverless inference (in beta), $1.87 and $4.68 on the training price list. At Baseten, Nebius, NVIDIA and Fireworks, I read no price this morning.

The Apache 2.0 license grants everyone a "perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable" right to reproduce, modify, redistribute and sublicense, patents included. The card adds a link to a lab usage policy, which states it binds anyone "accessing, downloading, or using" the files.

What we don't know yet

The external testers' report, if there is one. Prices at Baseten, Nebius, NVIDIA and Fireworks. And Europe: none of the four pricing pages read mentions a region or a country.

My opinion, labelled as such: Thinking Machines delivered on the date it said, and that deserves a mention. But the Tinker glass, the one that allowed access to be revoked, came down on Thursday evening. Apache 2.0 says "irrevocable". Their July text said "irreversible". For once, the contract and the lab say the same thing.

Words to remember

• Open weights: the model's files are published; anyone can download it and run it on their own machines.
• Parameter: one of the numbers tuned during training; their total gives the size of the model.
• Mixture of experts (MoE): a model split into sub-networks, only some of which compute for each token; here 8 experts out of 384.
• NVFP4: an NVIDIA format, introduced with its Blackwell generation of cards, that stores each number of the model on 4 bits.
• MXFP8: another format, storing each number on 8 bits; heavier.
• Token: the piece of text the model reads or writes at each step.

Source sheet

• Thinking Machines: post of 05/11/2026, update of 12/11/2026 · model card · "A Safe Path to Open Weights", 31/07/2026 · usage policy · Tinker pricing · "End‑to‑End 4‑Bit Training", 13/11/2026.
• Hugging Face: Inkling-2-Small (NVFP4) · Inkling-2-Small-MXFP8.
• Artificial Analysis: Inkling 2 Small page.
• Hosts: Together · Modal · Cloudflare.
• License: Apache 2.0.

Source passages

Show the 16 sources

Ref.	Quote (original language)	URL
S1↑1↑2↑3↑4↑5↑6+2	"Nov 5, 2026"; "Nov 12, 2026 Update: Inkling 2 Small's weights are now available for download on HuggingFace . It is also 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 ."; "It is a Mixture-of-Experts model with 594B total parameters, 17B active, and supports image inputs and variable effort."; "with a score of 31"; "Inkling 2 Small will initially be available through Tinker, with closed weights, following the approach we outlined in a safe path to open weights . We intend to make the full weights available under an Apache 2.0 license next week after the completion of external safety testing and red-teaming."; Safety section: "The release of Inkling 2 Small's weights is pending further safety evaluations by us and our partners."; "with more external testers, than any previous release"; "When tasks were difficult, early checkpoints of Inkling 2 Small were persistent in trying to succeed—in some cases fabricating results, ignoring instructions, and concealing what they had done."; "Midway through the main RL run, our asynchronous monitor flagged an instance where the model, failing at a task, worked around a network block through a third-party service it could reach using its search tools. We paused the run, hardened our infrastructure"; tables: "FORTRESS" (Scale AI), "HoneyBench v0.2-alpha: Reward Hacking (Goodhart Labs)", "Covert Action with No Goal (Apollo Research)"; "Training ran natively in NVFP4, the 4-bit floating-point format introduced with the NVIDIA Blackwell GPU architecture. We'll share more details on our NVFP4 recipe in a follow-up post soon."; "To improve consistency, we rely on externally reported evaluations for both internal and external models when applicable."; Try section: "Following the completion of safety testing, the model will be available on additional platforms in the coming weeks."	https://thinkingmachines.ai/news/inkling-2-small/
S2↑1↑2↑3↑4	API: "createdAt 2026-11-04T19:51:05.000Z"; "license:apache-2.0"; "gated False"; 25 safetensors files, 393,480,084,520 bytes; history: "2026-11-12T19:53:15.000Z upload model" (first upload), then 20:58:09Z, 21:35:28Z, 2026-11-13T03:42:44Z	https://huggingface.co/api/models/thinkingmachines/Inkling-2-Small?blobs=true; https://huggingface.co/api/models/thinkingmachines/Inkling-2-Small/commits/main
S3↑1↑2↑3	API: "createdAt 2026-11-04T19:51:29.000Z"; "license:apache-2.0"; 49 safetensors files, 622,803,514,224 bytes; main number type "F8_E4M3"; history: "2026-11-12T19:53:16.000Z upload model"	https://huggingface.co/api/models/thinkingmachines/Inkling-2-Small-MXFP8?blobs=true
S4↑1↑2↑3↑4↑5↑6	"license: apache-2.0"; "The main NVFP4 checkpoint requires a GPU cluster with at least 500GB of aggregated VRAM. This can be met with either of the following example configurations: 2x NVIDIA B300 GPUs 4x NVIDIA H200 GPUs"; "The alternative MXFP8 checkpoint requires a GPU cluster with at least 800GB of aggregated VRAM"; "We recommend the NVFP4 checkpoint for all use-cases"; "The MXFP8 has a very marginal 1–2 percentage points gain over the NVFP4 checkpoint over evals, but we do not think it is meaningful for any practical use-cases."; "each token is routed to 8 of 384 experts"; "including through multi-turn, open-ended external red-teaming"; "Across all areas, we concluded that Inkling 2 Small did not present risk of material uplift beyond what's already available in the open-weight ecosystem."; link "Acceptable Use"	https://huggingface.co/thinkingmachines/Inkling-2-Small
S5↑1↑2	"Thinking Machines Jul 31, 2026"; "Open models also carry real misuse risks, and release is irreversible."; "the provider can monitor usage, maintain guardrails, and revoke access"	https://thinkingmachines.ai/blog/a-safe-path-to-open-weights/
S6↑	"Date of release Nov 5, 2026"; "License Apache 2.0"; same hardware requirements as S4	https://thinkingmachines.ai/model-card/inkling-2-small/
S7↑1↑2	"Last updated July 15, 2026"; "By accessing, downloading, or using any Model Materials, you agree to be bound by this Model AUP"	https://thinkingmachines.ai/model-acceptable-use-policy
S8↑1↑2↑3↑4↑5	"# 4 / 28"; "31"; "Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index"; "Artificial Analysis Intelligence Index v5.0.1 incorporates 9 evaluations: AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode-AA 2.0, AA-Omniscience v1.1, GDP.pdf, HLE-Diamond, CritPt v2 (Hard Subset), and Terminal-Bench-Science 0.1"; "Large: >150B parameters"; "At 29.6 minutes per Intelligence Index task, Inkling 2 Small is notably slow compared with the median of 14.6 minutes"; page data: Inkling 2 Small "intelligenceIndex":30.6286781498958"; Kimi K3.1 Pro 46.45 (open weights); DeepSeek V4.1 Flash 1017 (Max) 32.59; GLM-5.3 (Max) 31.65; Claude Opus 5.5 (Max) 59.11; Gemini 3.8 Flash (High) 30.85; Inkling (Xhigh) 12.48; Claude Fable 5.2 (Max, Default Fallback) 64.56 (highest score on the page)	https://artificialanalysis.ai/models/inkling-2-small
S9↑1↑2↑3	"On Together AI it costs $0.30 per 1M input tokens and $1.00 per 1M output tokens."	https://www.together.ai/models/inkling-2-small
S10↑1↑2↑3	"Prompt $0.30 /MTok"; "Cached prompt $0.03 /MTok"; "Completion $1.00 /MTok"	https://modal.com/library/thinking-machines/inkling-2-small
S11↑1↑2↑3	"$0.30 per M input tokens, $1.00 per M output tokens, $0.10 per M cached input tokens"	https://developers.cloudflare.com/workers-ai/models/inkling-2-small/
S12↑1↑2↑3↑4	"All prices are per million tokens."; "Inkling 2 Small … thinkingmachines/Inkling-2-Small $0.30 $0.03 (cached) $1.00 $1.50" (Prefill, Sample, Train); "Inkling … thinkingmachines/Inkling $1.87 $0.374 (cached) $4.68 $5.61"; Serverless Inference (Beta): "Inkling thinkingmachines/Inkling:peft:262144:sampling-nvfp4 256K $1.00 $0.17 (cached) $4.05"	https://tinker-docs.thinkingmachines.ai/tinker/models/models_and_pricing/
S13↑	News: latest post "Introducing Inkling 2 Small Nov 5, 2026"; Blog: latest post "A Safe Path to Open Weights … Jul 31, 2026"; no external testing report for Inkling 2 Small	https://thinkingmachines.ai/news/; https://thinkingmachines.ai/blog/
S14↑1↑2	"each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work"; "3. Grant of Patent License"	https://www.apache.org/licenses/LICENSE-2.0
S15↑1↑2↑3↑4↑5↑6	Lab blog, post "End‑to‑End 4‑Bit Training", "Jack Cook in collaboration with others at Thinking Machines", "Nov 13, 2026", read on 14/11/2026 at 01:37:31 UTC (HTTP 200), re-read at 01:41:28 UTC (identical); absent from the blog index re-read on 13/11/2026 at 07:55:02 UTC, present in the index on 14/11/2026 at 01:36:59 UTC: "Inkling 2 Small is the largest model we are aware of that was trained end-to-end with a native 4-bit forward pass."; "We configure numerics in the same way from the start of pretraining through the final RL rollout."; "NVFP4 precision in the expert layers for forward passes, bolstered by a dedicated kernel."; "MXFP8 precision for backward passes with re-quantization and stochastic rounding"; "The backward passes that update the weights"; "In our largest jobs, this recipe provides us with an 8% wall-clock speedup in pretraining and a 20% speedup in RL over full MXFP8 training."; "In pretraining Inkling 2 Small it costs about 2% in compute efficiency, meaning the model reaches the loss it would have reached with about 2% fewer FLOPs if trained in MXFP8"; "We train on NVIDIA’s Grace Blackwell Ultra platform, the GB300"	https://thinkingmachines.ai/blog/end-to-end-4-bit-training/ ; https://thinkingmachines.ai/blog/
S16↑1↑2↑3↑4↑5	"# 4 / 29"; "31"; "At 28.1 minutes per Intelligence Index task, Inkling 2 Small is notably slow compared with the median of 14.3 minutes for comparable models"; "(based on Thinking Machines's API)"; page data: Inkling 2 Small "intelligenceIndex":30.6286781498958	https://artificialanalysis.ai/models/inkling-2-small