Inkling 2 Small benchmark:
31 points, 131 tokens a second, and a four-way Flash test

Artificial Analysis's current pages place Inkling 2 Small at 31 and Inkling Small at 13 on v5.0.1. The gap is 18 Index points. That is a striking step up on this scorecard, and Thinking Machines describes gains across agentic coding, vision, knowledge work and reasoning. But the arithmetic does not mean the new model is 138 per cent more capable, nor does it tell us that every team will see an 18-point lift in its own work. Composite points are not a universal unit of intelligence.

To see what 31 means in practice, place it beside the three models most buyers already weigh in the same fast, low-cost tier. DeepSeek Flash 1017 (DeepSeek-V4.1-Flash-1017) leads the lightweight board at 45 on the Intelligence Index. Anthropic's Claude Haiku 5.5 scores 34 at medium effort and climbs to 43 at maximum adaptive thinking. Google DeepMind's Gemini 3.8 Flash lands at 35. The jump from 13 to 31 brings Thinking Machines into direct contact with the frontier Flash class, even while leaving a clear gap to the composite leaders. (Artificial Analysis benchmark; Inkling Small score; AA Leaderboard)

The word "Small" could suggest a compact model. The published architecture is more interesting. Thinking Machines says Inkling 2 Small has 594 billion total parameters, with 17 billion active, in a 48-layer decoder-only Mixture-of-Experts model. Each token is routed to eight of 384 experts. DeepSeek Flash 1017 follows a remarkably similar sparse playbook: roughly 763 billion total parameters with about 16 billion active per token, paired with Context-Enhanced Dispersion and Cross-Stage Attention Acceleration.

In plain English, a router selects a subset of specialist networks for each token rather than activating every expert at once. That sparsity helps explain why total size and active parameters are both reported, and why Tinker prices MoE runs on active parameters. It does not, by itself, tell us the memory footprint, throughput, serving hardware or cost for a real deployment. One NeurIPS 2024 inference study found that MoE deployment can remain difficult because large total weights still demand substantial high-bandwidth memory and intricate expert communication. Think of the fox's trail as a routing path: 17 billion active parameters per token keeps per-step arithmetic brisk, while 594 billion total parameters hold the specialist breadth. (Thinking Machines model card; Tinker model listing; Huang et al., NeurIPS 2024; Switch Transformers, JMLR)

Sparse Mixture-of-Experts scale: total versus active parameters

Source: Artificial Analysis and Thinking Machines Lab, checked 8 November 2026.

One token through the MoE, and how rival architectures compare

If DeepSeek Flash 1017 and Claude Haiku 5.5 lead on composite Index score, streaming speed and per-task evaluation cost, why would an engineering team reach for Thinking Machines' model? The answer shows up in long-horizon software engineering and terminal benchmarks. Thinking Machines reports 86.3 per cent on SWE-bench Pro v2 (Hard) and 85.1 per cent on Terminal-Bench 2.1, alongside launch demonstrations of a Three.js city simulator, a Python PC emulator and iterative visual UI refinement from hand-drawn sketches.

Across the four models, each contender carves out a distinct territory. Inkling 2 Small is built to observe, compile, inspect and revise inside multi-step coding and terminal loops. Claude Haiku 5.5 pairs low-latency chat with class-leading GUI and desktop control (72.4 per cent on OSWorld 2.1, alongside 39.2 per cent on the stricter Terminal-Bench 4.0). Gemini 3.8 Flash brings tunable low, medium and high thinking budgets to massive multimodal repositories (DeepSWE v1.1). DeepSeek Flash 1017 delivers the highest raw reasoning-per-dollar throughput. Company-selected demos are not independent replications, so teams should treat that 86.3 per cent SWE-bench figure as an invitation to run a repo-level trial. (Thinking Machines announcement)

Agentic coding, terminal agency and composite scores

Source: Artificial Analysis and Thinking Machines Lab, checked 8 November 2026.

At announcement, Tinker provided inference and fine-tuning access, while Inkling 2 Small's weights remained closed under Tinker Terms of Use. Thinking Machines said it intended to publish full weights under Apache 2.0 after external safety testing and red-teaming. The company first said "next week"; its current Inkling page says "in the coming weeks" and still describes the model as closed-weight. Should those Apache 2.0 weights land, Thinking Machines' model and DeepSeek Flash 1017 would stand apart from proprietary endpoints like Haiku 5.5 and Gemini 3.8 Flash for self-hosted deployments.

Until then, pay close attention to hosted context limits and pricing cliffs. While the public model card states a maximum context of one million tokens, Tinker's live models.json currently lists 64K and 256K hosted options, with LoRA training priced at US$1.50 per million tokens on 64K and US$3.00 on 256K. Meanwhile, Claude Haiku 5.5 offers a one-million-token window, but its bargain US$0.10 input and US$0.50 output rate applies only up to 100K tokens; cross 100K and the price jumps fivefold to US$0.50 input and US$2.50 output. DeepSeek Flash 1017 ($0.30/$1.20, or $0.006 cache hit) and Gemini 3.8 Flash ($0.75/$3.75) hold flat rates across their full one-million-token windows. (Thinking Machines announcement; model card; Tinker model listing)

Hosted context ceilings and the 100K pricing cliff

Inkling 2 Small's current score of 31 marks a substantial rise over Inkling Small's 13 on Artificial Analysis's current Index, backed by a brisk 131-tokens-per-second output stream and a 594B-total, 17B-active sparse MoE design. Yet its upfront reasoning pause stretches Index evaluations to 15.7 minutes and US$1.18 per task; DeepSeek Flash 1017 and Claude Haiku 5.5 deliver higher composite scores at a fraction of the per-task spend; and the promised Apache 2.0 weights remain conditional on safety testing. The fair reading holds every piece of that picture at once.

Use the score of 31 as a shortlist signal, not a substitute for a field test. Pick the model whose latency, context economics and agency profile fit the work in front of you, and let a live repository decide which runner earns its keep.

Quick workload routing guide

Source: Artificial Analysis and Thinking Machines Lab, checked 8 November 2026.