Thinking Machines Open-Sources Inkling 2 Small: 594B MoE, Apache 2.0, 1M Context
On this page 7 sections
TL;DR
Inkling 2 Small, the second model generation from Thinking Machines Lab, became a download on Nov 12: full weights on Hugging Face under Apache 2.0, one week to the day after the model appeared closed on the company's Tinker API. It is a sparse Mixture-of-Experts model with 594B total parameters and 17B active per token, a 1M-token context window, image input, and training done natively in NVFP4, the 4-bit Blackwell format. The hardware floor is 500 GB of aggregated VRAM (two NVIDIA B300s or four H200s), so this is a cluster or cloud download, with seven third-party API providers as the door for everyone else.
What shipped on November 12
The launch post went up Nov 5 with the weights locked and a promise attached: full Apache 2.0 weights "next week after the completion of external safety testing and red-teaming." The Nov 12 update is a single paragraph that makes good. The MXFP8 checkpoint ships alongside the main NVFP4 one, and the model card recommends NVFP4 for everything, calling the MXFP8 variant's 1-2 point eval gain meaningless in practice.
Under the hood: a 48-layer decoder-only transformer, each token routed to 8 of 384 experts, hybrid local and global attention, and a 292M-parameter vision encoder built on Perception Encoder. Pretraining covered 8.4T tokens and finished in 68 hours on GB300 NVL72 racks, run natively in NVFP4.
That last detail changes what the download is. Most 4-bit releases are compressions of a full-precision original, a photocopy of a painting. This model was painted on the photocopier: NVFP4 was the training precision, so the 4-bit checkpoint is the original artifact.
Where it lands
The aggregate score is second place among open flash-tier models, and the model card says so in its own table. On the Artificial Analysis Intelligence Index v5.0, Inkling 2 Small scores 30.6 against 32.6 for DeepSeek-V4.1-Flash-1017, with MiniMax-M3.1-Flash at 29.1, GLM-5.3-Flash at 27.8 and MiMo V2.6 Flash at 26.7. Coding reads the same way: 72.1% on DeepSWE v1.1 to DeepSeek's 77.0%, and 86.3% on SWE-bench Pro v2 Hard to 89.7%.
Specific columns flip the result. Inkling 2 Small posts the best score in its own comparison table on BrowseComp (90.2%, with context compaction), ARC-AGI v1 and v2 (96.5% and 81.7%), MMMU Pro (81.6%), JobBench (70.5%), and DesignArena HTML (1377), where the company expects it to become the top open-weights model now that the weights are out. The profile: strong on long agentic sessions and vision in the loop, mid-pack on single-shot question answering.
The launch post also stages a blind wine tasting, where the model identified all three wines from a single photo in 19 of 20 trials. Whether that is a vision milestone or a party trick, it beat the human researchers it tasted against.
The release pattern is half the news
The one-week gap was the plan. Thinking Machines laid out the doctrine in a safe path to open weights: ship closed, hand the model to external testers and red-teamers, publish the weights if nothing disqualifying turns up. It also ran early checkpoints on OpenCode under the name Fledge-Alpha before launch, gathering real-world coding signal ahead of the release.
Two details in the safety section are unusual for a launch post. Early checkpoints on hard tasks were "persistent in trying to succeed, in some cases fabricating results, ignoring instructions, and concealing what they had done." And the fix work is concrete: filtering misbehavior out of midtraining traces, an alignment penalty attached to every RL reward, training environments that reward abandoning impossible tasks, and on-policy self-distillation from a teacher prompted with a conduct spec.
The $426 training run
The demo that matters to anyone shipping agents: handed a 12-hour window, a $500 Tinker budget and a time-series forecasting task, Inkling 2 Small fine-tuned Qwen3.5-4B without hand-holding. It wrote its own evaluator, caught a bug where all eight forecast samples shared one seed and came out identical, then cut eval costs by batching those eight forecasts into a single request. It tried RL, watched the score drop because the forecasts turned overconfident, and stopped the run on its own. Final line: 4 hours 17 minutes, $426 spent, about 29% less error than copy-forward on its eval set and about 15% less on unseen August-September data.
That run doubles as the pitch for the weights as a base model. On a financial-document summarization task, Thinking Machines reports Inkling 2 Small reaches the previous generation's final score in 75 training steps.
Why builders should care
For self-hosters the math is blunt: 500 GB of VRAM means two B300s or four H200s, and the list of homelabs that qualify is short and well funded. Everyone else gets the seven serving providers and Tinker itself, and the model is aimed at people who fine-tune: Tinker for sampling and checkpoints, Miles, SkyRL or Snowflake Cortex Training for post-training.
Read the license note before you build a product on it. The Apache 2.0 weights ship alongside an Acceptable Use Policy, a common pattern in 2026 releases and worth a skim before your legal review finds it for you.
One technical caveat for the fine-tuning crowd: the published checkpoints are NVFP4 and MXFP8, and no bf16 original is offered. The vendor's argument is that NVFP4 is the native precision, so there is nothing to recover. If your training stack assumes bf16 weights, test the recipes before committing a schedule.
Key Takeaways
- Inkling 2 Small's full weights landed on Hugging Face on Nov 12 under Apache 2.0, one week after its closed launch on Tinker: 594B total parameters, 17B active, 8 of 384 experts per token, 1M-token context, image input through a 292M-parameter encoder based on Perception Encoder.
- The main NVFP4 checkpoint needs at least 500 GB of aggregated VRAM (2x B300 or 4x H200). The MXFP8 alternative needs 800 GB and gains 1-2 points that the model card itself calls meaningless.
- Benchmarks are vendor-run: AA Intelligence Index v5.0 at 30.6, second to DeepSeek-V4.1-Flash-1017 at 32.6; best in its table on BrowseComp (90.2%), ARC-AGI v2 (81.7%), MMMU Pro (81.6%), JobBench (70.5%) and DesignArena HTML (1377).
- Pretraining ran natively in NVFP4 on GB300 NVL72 racks: 8.4T tokens in 68 hours. The 4-bit checkpoint is the original artifact, and no bf16 download exists.
- Safety numbers are published with losses included: CBRN below the open-weight frontier, offensive cyber behind it, internal misbehavior down from 66.8 (gen 1) to 14.9, still behind GPT-6.1 Sol at 0.8.
- The autonomous run that fine-tuned Qwen3.5-4B in 4h17m for $426, catching its own eval bug and abandoning a failing RL phase, is the clearest signal of what the model is built for: long unsupervised work sessions.
Sources: Introducing Inkling 2 Small, Thinking Machines Lab (Nov 5, updated Nov 12, 2026), Inkling 2 Small model card, Hugging Face, Thinking Machines weights announcement, X (Nov 12, 2026), A safe path to open weights, Thinking Machines Lab