Inkling 2 Small
Inkling 2 Small (thinkingmachines/inkling-2-small) is Thinking Machines Lab's open-weight, sparse Mixture-of-Experts reasoning model — 594B total / 17B active parameters — released 5 Nov 2026 with a 1M-token context window, text-and-image input, and variable-effort reasoning RL'd on long-horizon coding across multiple agent harnesses. It posts frontier-adjacent SWE-bench Verified (82.2%), HLE no-tools (44.1%) and GPQA Diamond (90.9%) numbers at roughly $0.30/1M input and ~105 tok/s throughput, while collapsing sharply on long-horizon terminal, SRE and program-synthesis tasks.
Empirical Evaluation Results
Architectural Profile & Capabilities
MODEL IDENTITY & RELEASE Inkling 2 Small is the second-generation, size-optimized member of Thinking Machines Lab's Inkling family, with OpenRouter slug thinkingmachines/inkling-2-small and a public release date of 5 November 2026 (Hugging Face open-weights drop followed within days). It is explicitly NOT the larger Inkling 2 flagship (which reports, e.g., a higher 87.7% figure on Vibe Code Bench), and it is NOT the earlier Inkling Small / Inkling 1 generation. Every score reported in this dossier was published against the Inkling 2 Small name/listing specifically; no sibling or flagship proxies were used. ARCHITECTURE Sparse Mixture-of-Experts (MoE) autoregressive transformer: 594B total parameters with 17B active parameters per token — a ~2.9% activation ratio, which is why it can serve at ~104–256 tokens/sec despite a half-trillion-parameter weight footprint. Context window is 1,048,576 tokens (1M). The official announcement and model card describe text-and-image inputs with text-only outputs; OpenRouter's normalized 'text' modality flag reflects output modality only, not a text-only input restriction. Audio is not supported. Attention specifics (MLA vs GQA vs sliding-window/hybrid sparse attention) are not publicly established for this variant and are deliberately not asserted here. Tool use is supported natively, and reasoning depth is user-controllable via variable reasoning effort. TRAINING PARADIGM Thinking Machines describes post-training as reinforcement learning on long-horizon coding tasks executed across multiple agent harnesses — i.e., the model was optimized against real harnesses (OpenHands-style scaffolding, terminal environments) rather than single-turn SWE patches alone. Stated target domains: agentic coding, autonomous research, agentic knowledge work, and retrieval-augmented generation. No peer-reviewed technical report detailing SFT stages, reward-model design, or reasoning-trace format has been published for this variant. EMPIRICAL PROFILE — WHAT IT IS ACTUALLY GOOD AT The published record shows an unusually bimodal model. On short-to-medium-horizon, high-signal academic and single-harness evaluations it performs like a much larger system: SWE-bench Verified 82.2%, DeepSWE v1.1 72.1%, GPQA Diamond 90.9%, HLE (no tools) 44.1%, HLE-Diamond 28.7%, BrowseComp Search 90.2%, Terminal-Bench 2.1 81.3%, plus strong professional-judgment Elo scores (GDPval-AA v2.1 1630, AA Briefcase Presentation 1564, AA Briefcase v1.1 1527, Design Arena Frontend 1294). On long-horizon, environment-fragile, or stateful production-simulation tasks the same model falls off a cliff: Terminal-Bench 4.0 13.1%, Terminal-Bench Science 5.7%, FrontierCode 2.0 25.4%, SRE Bench 1.9%, ProgramBench 0.5%. Notably it scores 81.5% on Vibe Code Bench v1.1 (web-app construction from scratch) but only 0.5% on ProgramBench — evidence that it is strong at generating plausible application code and weak at constrained, verifier-graded program synthesis. Aggregated intelligence reads low (AA Intelligence Index v5.0 = 30.6; AA-Omniscience Index 16.1; Slop Index 25.4), consistent with a model whose strength is narrow and bench-specific rather than broadly calibrated. ECONOMICS & SERVING Benchmark Heaven telemetry: $0.30/1M input tokens, $1.00/1M output tokens. Host-dependent pricing and quantization vary widely — the OpenRouter listing for the same variant has appeared at $0.15/1M input and $0.50/1M output with $0.03/1M cache reads, alongside a free tier. Provider-level measurements: TTFT 0.57–1.16s and throughput 81–256 tok/s (Cloudflare ~256, Modal ~186, Together ~173, DeepInfra ~126, Nebius ~123, Prime Intellect ~81); the Thinking Machines-hosted free endpoint measured ~1.13s latency and ~104 tok/s, matching Benchmark Heaven's ~105 tok/s.\n\nPOSITIONING (OCTOBER 2026 FRONTIER)\nInkling 2 Small is not a frontier-class generalist. Against GPT-6 Astra / GPT-5.6 Sol, Claude Opus 5 / Sonnet 5, Gemini 3.8 Pro / Flash, DeepSeek V4 Pro, Grok 4.6, Qwen 3.8 and Kimi K3, its case rests on being open-weight, cheap, extremely fast, 1M-context, and unusually strong on single-harness SWE and STEM-reasoning benchmarks relative to its active-parameter budget. Buyers should treat the high scores as harness-shaped rather than capability-general.
594B total parameters with 17B active parameters per token (~2.9% activation ratio). 1,048,576-token (1M) context window. Text-and-image input, text-only output (no audio). Native tool use with user-controllable variable reasoning effort. Post-trained with reinforcement learning on long-horizon coding tasks across multiple agent harnesses; targeted at agentic coding, autonomous research, agentic knowledge work and RAG. Released as open weights on Hugging Face (thinkingmachines/Inkling-2-Small) on 5 Nov 2026, alongside an official model card and release announcement. Attention mechanism (MLA vs GQA vs sliding-window/hybrid) is not publicly disclosed for this variant and is not asserted. Reported at $0.30/1M input and $1.00/1M output tokens in Benchmark Heaven telemetry, with host-dependent pricing as low as $0.15/$0.50 on some OpenRouter routes.
Recommended Workloads & Primary Use Cases
- •Severe long-horizon terminal and infrastructure regression: Terminal-Bench 4.0 = 13.1% and Terminal-Bench Science = 5.7% versus Terminal-Bench 2.1 = 81.3%. The model's agentic competence is harness- and horizon-sensitive; multi-step shells, stateful ops and SRE-style incident work (SRE Bench 1.9%) should not be delegated autonomously.
- •Weak verifier-graded program synthesis: ProgramBench = 0.5% while Vibe Code Bench v1.1 = 81.5%. It produces plausible application code but fails constrained, spec-exact program construction — enforce compiled test gates rather than accepting generative output.
- •Low calibration and knowledge-honesty indices (AA-Omniscience Index 16.1, Slop Index 25.4, AA Intelligence Index v5.0 = 30.6) imply elevated hallucination and verbosity risk on open-ended factual work; pair with retrieval grounding and answer-abstention prompting rather than trusting internal knowledge.
- •1M-token context is nominal, not uniformly usable: KV-cache pressure, TTFT growth and recall degradation on needle-in-haystack tasks are expected, and no MRCR v2 long-context score has been published for this variant — validate retrieval recall above ~200K tokens before relying on full-window prompts.
- •Reasoning-effort and would-be multimodal settings are non-deterministic across hosts: pricing, quantization and TTFT vary by provider ($0.15-$0.30 input, 81-256 tok/s), and OpenRouter's 'text' modality tag contradicts the official text+image input spec, so pin a single provider and verify image-input support per route before production rollout.
- •Published scores cluster around Vals AI and Artificial Analysis harnesses with different scaffolds (e.g., Terminal-Bench 4.0 reported as 13.1% by AA and 12.6% by Vals AI); the wide spread between same-named benchmarks confirms harness sensitivity — reproduce results in your own harness before capacity planning.
Calls to Inkling 2 Small are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by Thinking Machines.
More Models from Thinking Machines
Compare other engines available in this laboratory.
Inkling is Thinking Machines Lab's first open-weight foundation model — a 975B-parameter sparse Mixture-of-Experts multimodal transformer (41B active parameters, released July 15, 2026) that reasons over text, images, and audio with a controllable "thinking effort" dial. Built for general-purpose reasoning, coding, and agentic/tool-use workloads, it delivers broad, balanced capability with strong open-weights standing (e.g., #1 US open-weights model on Artificial Analysis) at ~210 tokens/sec hosted throughput.
Inkling Small is a high-performance foundation model engineered by Thinking Machines, accessible with Zero Data Retention on ARMES.