R067 · Infrastructure for agents at scale › Frontier model releases

Inkling 2 Small: an open-weight 594B MoE built for agent loops and fine-tuning

Read the original

model-releasesopen-weightsagent-harnessreinforcement-learning

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

ArchitectureSparse mixture-of-experts, 594B total / 17B active parameters
InputsText and images (292M-parameter vision encoder); text out; no audio
ContextUp to 1M tokens; trained at 256K, with compaction for longer agent runs
ReasoningVariable effort
WeightsApache 2.0. Released closed on 5 Nov via Tinker, then open on Hugging Face on 12 Nov after external safety testing and red-teaming
ServingTinker, NVIDIA, Fireworks, Together, Baseten, Nebius, Modal, Cloudflare; vLLM, SGLang, Unsloth, TokenSpeed; Ollama Cloud with zero data retention
Post-trainingTinker, Miles, SkyRL, Snowflake Cortex Training

How it was trained

Claims worth checking

Stated limitations

What it means for agent builders

  1. 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.

  2. 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.

  3. 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.

  4. Skill routing is a known weak spot. If your agent exposes many skills, measure routing accuracy before switching models.

Review checklist before adopting

  1. Re-run your own agent eval (with your harness and compaction settings), not the vendor's.

  2. Measure tool and skill selection with your real catalog size.

  3. Compare at least two providers for tool-call error rate and latency.

  4. Check the license on derived checkpoints and the model card's use restrictions.

  5. 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.

Connections

See R067 on the knowledge map

Part of the Research notes on nowucca.com. Figures and claims are the source's own unless the summary says otherwise.