Sources: Ollama's LinkedIn post announcing the model on Ollama Cloud (2026-11-13); Thinking Machines Lab, Introducing Inkling 2 Small (5 Nov 2026, updated 12 Nov) and its model card; AI/TLDR's release summary. All benchmark numbers below are the vendor's own.
What shipped
| Architecture | Sparse mixture-of-experts, 594B total / 17B active parameters |
| Inputs | Text and images (292M-parameter vision encoder); text out; no audio |
| Context | Up to 1M tokens; trained at 256K, with compaction for longer agent runs |
| Reasoning | Variable effort |
| Weights | Apache 2.0. Released closed on 5 Nov via Tinker, then open on Hugging Face on 12 Nov after external safety testing and red-teaming |
| Serving | Tinker, NVIDIA, Fireworks, Together, Baseten, Nebius, Modal, Cloudflare; vLLM, SGLang, Unsloth, TokenSpeed; Ollama Cloud with zero data retention |
| Post-training | Tinker, Miles, SkyRL, Snowflake Cortex Training |
How it was trained
Pre-training in 4-bit. The first phase covered 8.4T tokens of text, code and images in 68 hours on GB300 NVL72 racks, natively in NVFP4. A recipe post is promised. If it holds up, it is a notable data point that frontier-scale pre-training can run in FP4 from the start.
One large RL run of 1,075 steps and over 17M rollouts, over 9 days on 2,048 GPUs.
Coding RL across harnesses. Hundreds of thousands of long-horizon coding tasks ran across several agent harnesses and context-compaction strategies. The reward went beyond unit tests: an agentic grader also judged how the model approached the problem. Early checkpoints were tested in the wild on OpenCode under the codename Fledge-Alpha.
Claims worth checking
Overall: 31 on the Artificial Analysis Intelligence Index, well above Thinking Machines' earlier models.
Coding: 72.1% on DeepSWE v1.1, 86.3% on SWE-Bench Pro v2 (Hard), 81.6% on Vibe Code Bench v1.1.
Held-out code compression: lowest bits-per-byte on the vendor's internal code set (cutoff June 2026) among the open-weight models compared, including Kimi K3.1 Pro and DeepSeek V4.
Fine-tunability: learns faster than the earlier Inkling models and Qwen, GLM and DeepSeek peers on the same data and algorithm. On a financial-summarization task, it matched the final score of its predecessor in about 75 steps.
Safety: an internal misbehavior rate of 14.9, down from 66.8 for the first Inkling. The vendor says it still trails the leading closed model on its alignment evaluations.
Stated limitations
Wordy, and doesn't always pick the right tool, especially in codebases with many skills. This matters for anyone running large skill catalogs.
Less tuned for chat: weaker formatting, overuse of slashes, and overreaction to mental-health and safety queries.
No audio input, unlike the earlier Inkling-Small.
What it means for agent builders
A US-built, Apache 2.0 base at the frontier-open tier. For teams that need to fine-tune and self-host under permissive terms, this is the main reason to look.
Built to be fine-tuned. The vendor optimized for improvement under post-training, not only out-of-the-box scores. If you plan to specialize a model, compare learning curves on your own data, not leaderboard rank.
Expect provider variance. The same weights now run on a dozen providers and runtimes. Tool-call validity and long-context behavior can differ by provider and quantization, so pin a provider and test it.
Skill routing is a known weak spot. If your agent exposes many skills, measure routing accuracy before switching models.
Review checklist before adopting
Re-run your own agent eval (with your harness and compaction settings), not the vendor's.
Measure tool and skill selection with your real catalog size.
Compare at least two providers for tool-call error rate and latency.
Check the license on derived checkpoints and the model card's use restrictions.
If fine-tuning, budget a small learning-curve comparison against your current base model.
Bottom line
A serious open-weight release aimed squarely at agentic coding and customization, with an unusually candid limitations section. All performance numbers are self-reported. The 4-bit pre-training recipe and the fine-tunability claim are the most interesting parts to watch for independent confirmation.