Inkling 2 Small review, pricing and limits

Inkling 2 Small from Thinking Machines Lab reports 93.0% on GPQA Diamond and 82.2% on SWE-bench Verified in the vendor's own results. It suits teams fine-tuning an agent on Tinker, not buyers who need audio input, fast replies or a polished chat voice.

Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index, according to Artificial Analysis. Thinking Machines Lab released the multimodal mixture-of-experts model on 5 November 2026 with 594 billion total and 17 billion active parameters, text and image input, and text output.

Provider: Thinking Machines Lab · Family: Inkling

More about Thinking Machines Lab on Hokai

Context window: 1,000,000 tokens

Input modalities: text, image · Output: text

About Inkling 2 Small

Inkling 2 Small is the second-generation model from Thinking Machines Lab, released on 5 November 2026. It follows Inkling, released with open weights on 15 July 2026, and the smaller Inkling-Small, previewed the same day, and the company describes it as one rung on a ladder of model sizes with more to come. Thinking Machines positions it as a general-purpose model for agentic and tool-use systems, coding assistants, chatbots and retrieval-augmented generation, and says it was trained to be a strong base for customization on its Tinker platform.

Every score below is vendor-reported from the company's own comparison table unless it names Artificial Analysis, an independent evaluator that gives the model an Intelligence Index of 31 on version 5.0.1. The model card lists 82.2% on SWE-bench Verified, 93.0% on GPQA Diamond and 81.6% on MMMU-Pro. Its strongest rows are applied work: Finance Agent v2.1 at 67.5%, JobBench at 70.5% and EnterpriseOps-Gym at 46.8% are each the highest in the set it compares against. Pure coding and terminal rows are weaker. On Terminal Bench v4.0 it scores 13.1% against 32.8% for GLM-5.3 Flash, and DeepSeek V4.1 Flash 1017 leads it on DeepSWE v1.1 at 77.0%. Claude Haiku 5.5, a closed model, scores 36.5 on the version 5.0 index in the same table.

The context window is stated as up to 1M tokens by both the model card and Artificial Analysis, and the company capped its benchmark trajectories at the same length. Tinker, the only place to use the model today, serves a 64K standard context and a 256K extended context, so a Tinker customer cannot yet call the full 1M. No maximum output length is published. On long-context reasoning the card reports 83.3% on AA-LCR and 70.6% on LongBench v2.

Input is text and images, with output as text; there is no audio support, which the first-generation Inkling had. Tool use is a design focus, and the card reports 85.4% on the full MCP Atlas set and 85.0% on the public set. Vision results include 83.8% on ScreenSpot Pro, rising to 91.5% when the model may use Python, and 64.7% on PerceptionBench. The vendor shows the model building websites, games and 3D scenes by looking at screenshots of its own output and revising.

The vendor's own autoresearch demo gives a sense of cost at scale: the model trained a Qwen3.5-4B forecasting model on Tinker for 4.5 hours and $426 of Tinker spend. Artificial Analysis puts the cost of running its Intelligence Index at $1.18 per task. Pricing details are in the FAQ and the pricing table.

Access today runs through Tinker: the Tinker Playground for inference, the Tinker API for sampling and fine-tuning, and a serverless inference beta behind a waitlist. Thinking Machines says more platforms follow once safety testing finishes. The weights are closed at launch. The company says it intends to publish them under Apache 2.0, and the Hugging Face repository was not public on 9 November 2026.

Safety testing ran before release, both internal and external. On FORTRESS, Scale AI's private adversarial benchmark, the average risk score is 16.2 (lower is better) with an overrefusal score of 20.4. Apollo Research's covert-action evaluation with a goal scored 0. Its internal misbehavior measure is 14.9, and the company says it still trails closed models such as GPT-6.1 Sol, at 0.8, on alignment evaluations. For cyber, it reports matching the open-weight frontier on CyberBench v1.1 and PatchEval while staying behind that frontier on offensive agentic benchmarks.

It suits teams that want to fine-tune a mid-sized multimodal model, or run agents over spreadsheets, filings and documents, where the vendor reports its best results. It fits poorly with audio products, latency-sensitive chat or buyers who need a polished conversational voice. Thinking Machines says it is not the most succinct communicator, can pick the wrong tools in codebases with many skills, and in chat shows worse content organization and overreacts to mental-health queries. For peak coding, compare Kimi K3 first.

Pretraining covered 8.4 trillion tokens of text, code and images, followed by 3.8 trillion tokens of newer data at 512K context, about 320 billion tokens of reasoning and agent trajectories at 1M context, and about 34 billion tokens of top-quality agent trajectories. One joint reinforcement-learning run then used 1,075 steps and over 17 million rollouts across 9 days on 2,048 GPUs, replacing the separate reasoning and agentic runs used for the first generation. The later-phase trajectories were mostly sampled from earlier models including Kimi K3 and the company's own. The training cutoff is not disclosed.

Pricing

Sampling on Tinker costs $0.30 per million prefill tokens, $0.03 per million cached prefill tokens and $1.00 per million output tokens at either context size. Fine-tuning is billed on its own training rate, higher for the extended context. The serverless beta uses the same sampling rates.

The Tinker pricing page lists no free tier.

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• Sparse mixture of experts: Routes each token to 8 of 384 experts across 48 layers, leaving a small share of its 594B parameters active per token.
• 1M-token context window: The model card and Artificial Analysis both list up to 1M tokens, and the vendor ran its evaluations with a 1M-token trajectory limit.
• Image input with variable effort: Reads images from 40 to 4,096 pixels per side and lets callers set reasoning effort; the vendor's benchmarks use effort 0.99.
• Built for fine-tuning on Tinker: Thinking Machines runs the training infrastructure, and the vendor reports the model reaches the first generation's final score in 75 steps on a financial-summarization task.
• Tool use over MCP: Scores 85.4% on the full MCP Atlas set, second-highest in the vendor's table, and works with the tml-renderers package for tool-call formatting.

Pros

• Artificial Analysis rates its intelligence well above the 23 median it reports for comparable models.
• A 90% discount applies to cached input tokens on Tinker.
• Ships with a published model card, a full benchmark table and named external safety results.

Cons

• Artificial Analysis calls it notably slow, at 15.5 minutes per Intelligence Index task against a 7.2-minute median.
• Trails GLM-5.3 Flash and DeepSeek V4.1 Flash 1017 on several coding and terminal rows in the vendor's own table.
• Available only on Tinker for now, with no cloud-marketplace listing yet.

Benchmarks

• MMMU-Pro: 81.6% vendor-reported · 09 Nov 2026 — College-level questions that need reading images and diagrams, % correct.
• GPQA Diamond: 93% vendor-reported · 09 Nov 2026 — PhD-level science questions that are hard to search for, % correct.
• SWE-bench Verified: 82.2% vendor-reported · 09 Nov 2026 — Real GitHub issues fixed end to end, % solved.
• AA Intelligence Index: 31 cited: Artificial Analysis · 09 Nov 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
• Output speed: 161 tok/s cited: Artificial Analysis · 09 Nov 2026 — Median tokens written per second as measured by Artificial Analysis.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit Inkling 2 Small Official Page