Is Inkling 2 Small open source?

Not yet. It launched with closed weights on Tinker. Thinking Machines says it intends to release the full weights under Apache 2.0 after external safety testing and red-teaming, which its announcement put at the following week; Artificial Analysis lists the release date as unconfirmed. As of November 9, 2026 there is no public Inkling 2 Small repository on Hugging Face. Tinker fine-tune checkpoints become downloadable when the weights ship.

How much does Inkling 2 Small cost?

On Tinker's pricing page (November 9, 2026): $0.30 per million prefill (input) tokens, $0.03 cached, $1.00 per million sampled (output) tokens, and $1.50 per million training tokens at 64K context or $3.00 for the 256K-context variant. Artificial Analysis measured about 150,000 output tokens per Intelligence Index task, so a task costs about $1.18, roughly six times Inkling Small but well below Gemini 3.8 Flash (high) at $7.77.

Is Inkling 2 Small good for coding?

Mixed. Thinking Machines reports 72.1% on DeepSWE v1.1 and 25.4% on FrontierCode v2.0, second among the open-weight models it compared. Artificial Analysis found it ahead of Claude Sonnet 5.5 (medium) on FrontierCode 2.0 (25% vs 11%) but behind on Terminal-Bench 4.0 (13% vs 30%). It was trained on long-horizon tasks across several agent harnesses and tested in OpenCode under the name Fledge-Alpha.

Quick Answer

What Is Inkling 2 Small? Price, Benchmarks and Weights

Q: What is Inkling 2 Small?

Inkling 2 Small is Thinking Machines Lab's second-generation model, announced November 5, 2026: a Mixture-of-Experts model with 594 billion total parameters and 17 billion active, image input, text output and variable reasoning effort. It scores 31 on the Artificial Analysis Intelligence Index v5.0.1, matching Gemini 3.8 Flash (high) and 18 points above Inkling Small. It runs on Thinking Machines' Tinker platform for sampling and fine-tuning.

Published: November 9, 2026

The short answer

Inkling 2 Small is Thinking Machines Lab’s November 5, 2026 model: a 594B-parameter MoE (17B active) that scores 31 on the Artificial Analysis Intelligence Index, costs $0.30 / $1.00 per million tokens on Tinker, and is promised as Apache 2.0 open weights after safety testing. As of November 9, 2026, those weights are not yet public.

Key facts

	Inkling 2 Small
Developer	Thinking Machines Lab
Announced	November 5, 2026
Architecture	MoE, 594B total / 17B active, image input
Where to run it	Tinker and Tinker Playground (sampling and fine-tuning)
Price (Tinker)	$0.30 input · $0.03 cached · $1.00 output per 1M tokens; training $1.50 (64K) or $3.00 (256K) per 1M
Context on Tinker	64K and 256K variants
AA Intelligence Index v5.0.1	31 (Gemini 3.8 Flash high: 31; Inkling Small: 13)
Output tokens per AA task	~150K (Sonnet 5.5 medium ~31K)
Cost per AA task	$1.18 (Gemini 3.8 Flash high $7.77, Sonnet 5.5 medium $10.02)
Weights	Closed at launch; Apache 2.0 release planned, date unconfirmed

What changed from Inkling Small

Thinking Machines credits better pre-training efficiency and one large, unified reinforcement-learning run across coding, vision, knowledge work and reasoning. Artificial Analysis measured the jump: AutomationBench-AA from 5% to 46%, FrontierCode 2.0 from 5% to 25%, and hallucination rate on AA-Omniscience down from 61% to 27% — mostly because the model now declines more questions (attempt rate 77% → 53%) rather than knowing much more (accuracy 32% → 34%).

The model is trained to check its own work visually: Thinking Machines demonstrates it building websites, games and 3D scenes and correcting them from screenshots, and says it is set to top DesignArena among open-weight models once the weights ship.

Coding and agent use

Vendor numbers for the coding table: DeepSWE v1.1 72.1%, FrontierCode v2.0 25.4%, SWE-Bench Pro v2 (Hard) 86.3% — strong for its size but behind DeepSeek-V4.1-Flash-1017 on DeepSWE (77.0%). Independent results are less flattering on terminal work: 13% on Terminal-Bench 4.0 against 30% for Claude Sonnet 5.5 (medium). Thinking Machines also showcases “autoresearch”: in one run the model post-trained Qwen3.5-4B on Tinker for 4.5 hours and $426 of compute and more than doubled its forecasting score.

The catch is verbosity. At ~150K output tokens per Intelligence Index task, the cheap per-token price does not mean a cheap task. Measure cost per finished job, not per million tokens.

How it compares on price

Its $0.30 / $1.00 list price sits near other efficient open-weight MoEs: Reflection Beam 1.0 from $0.25 / $1.00 on its cheapest host, DeepSeek V4.1 Flash $0.15 / $0.60 off-peak, and well below Kimi K3.1 Pro at $2 / $10 (see our current API prices). The difference is the platform: today Inkling 2 Small runs only on Tinker, which is built for fine-tuning, so its natural buyer is a team that wants to post-train an open model and later download the checkpoint.

Should you use it now?

Yes, if you fine-tune on Tinker: you can train on Inkling 2 Small today and download checkpoints when the weights land.
Wait, if you need it on your own GPUs or a general inference host — nothing is downloadable yet.
Look elsewhere, if you need the strongest terminal coding agent for the money; see Tinker vs OpenAI fine-tuning vs Together AI for the fine-tuning angle and our explainer on the first Inkling model.

Last verified: November 9, 2026.