Inkling 2 Mini
Inkling 2 Mini is Thinking Machines Lab's open-weight (Apache 2.0) looped-transformer MoE released 20 Nov 2026: 26.4B total parameters with 3.4B active per token (~5.4B effective activations once middle-layer looping is counted), a 1M-token context window, and roughly 280 tokens/sec measured throughput at $0.075/$0.30 per 1M tokens. It is an efficiency play — near-flagship-tier coding and agentic-tool scores (SWE-bench Verified 79.2%, SWE-bench Pro v2 Hard 77.9%, MCP Atlas Public 82.7%) at roughly one-hundredth of frontier per-token cost — but a weak generalist on terminal/OS-execution work, deep scientific reasoning, and factual recall (AA-Omniscience Index −8.6).
Empirical Evaluation Results
Architectural Profile & Capabilities
ARCHITECTURE & TRAINING Inkling 2 Mini is a sparse Mixture-of-Experts decoder-only Transformer with a looped (recurrent-depth) core, announced by Thinking Machines Lab on 20 November 2026 alongside an official model card, a Hugging Face open-weights repository (Apache 2.0) and an OpenRouter endpoint (thinkingmachines/inkling-2-mini). The stored network is 24 decoder layers organised into four blocks of six. The two middle six-layer blocks (12 layers total) are executed three times with shared weights, while the first and final blocks execute once — yielding 48 layer applications per generated token from only 24 layers of stored weights. Each MoE layer routes each token to 8 of 128 experts. Official documentation reports 26.4B total parameters, 3.4B active parameters per token and approximately 5.4B effective weight activations per token once the repeated loop passes are counted. Attention is a hybrid of local and global layers; the vendor materials do not confirm GQA or MLA, so neither should be assumed. A vision Transformer encodes images into a shared hidden space that the decoder processes jointly with text, making the model natively multimodal on input (Thinking Machines describes it as a compact vision-language model; OpenRouter currently exposes it as a text endpoint). Supported image dimensions are 40px–4096px per side. The context window is 1,048,576 tokens (~1M). Tokenizer details, data mixture, data cutoff, compute budget and the RL/preference-tuning recipe are not disclosed in the public model card. POSITIONING IN THE OCTOBER–NOVEMBER 2026 LANDSCAPE Inkling 2 Mini is deliberately not a frontier-class reasoning model. On the Artificial Analysis Intelligence Index v5.0 it scores 20.1 — tied with Qwen3.8-27B and an order of magnitude below GPT-6 Astra, Claude Opus 5, Gemini 3.8 Pro and Kimi K3. Its real competitive claim is the score-per-dollar and score-per-gigabyte curve: Terminal-Bench 4.0 is only 5.6% and Terminal-Bench Science 1.4%, yet SWE-bench Verified hits 79.2% and SWE-bench Pro v2 (Hard) 77.9%, both ahead of Qwen3.8-27B (73.5% / 73.8%), while MCP Atlas Public (82.7%), DeepSearchQA with compaction (85.3%) and BrowseComp with compaction (79.0%) place it at the top of the sub-30B open-weight class for retrieval-augmented and tool-orchestration agents. Thinking Machines states it leads open-weight models of similar size on BrowseComp, DeepSearchQA and MCP Atlas. Its ~20 GB quantised footprint means it can be self-hosted on a single 24–32 GB GPU or a high-RAM workstation, which is the decisive operational fact for air-gapped and cost-capped deployments. MEASURED THROUGHPUT Benchmark Heaven telemetry records ~280 output tokens/sec, i.e. genuinely high-throughput generation for a 26.4B-total model, enabled by the 3.4–5.4B active-parameter budget rather than by any speculative-decoding trick. That figure reflects short-to-medium-context server inference; long prefill on 1M-token documents is dominated by attention cost and local GGUF/quantised deployments fall far below it. STRENGTHS (1) Repository-level code editing and patch generation at commodity pricing. (2) MCP/API tool orchestration: the model was tuned toward agentic search-compaction loops, and its BrowseComp/DeepSearchQA/MCP Atlas numbers depend on harness-level compaction of tool returns — the model itself is not the only variable. (3) Cheap structured extraction from long PDFs and business documents (GDP PDF 16.8%, AnalystAgent 46.3%). (4) Design/frontend generation (Design Arena Frontend 1281 Elo, 3D & GameDev 1304 Elo) is unusually strong for its price class. (5) Abstention discipline is above average for its intelligence tier (AA-Omniscience Non-Hallucination 63.4%), even though its raw knowledge index is negative. WEAKNESSES Terminal/OS automation is close to unusable (Terminal-Bench 4.0 5.6%, Terminal-Bench Science 1.4%); hard science reasoning is poor for the price (CritPt v2 Hard 17.0%, HLE-Diamond 15.2%); the negative AA-Omniscience Index (−8.6) signals that obscure factual recall is unreliable and that the model will sometimes answer instead of abstaining. The looped design also means reasoning effort is coupled to compute depth: reported GPQA Diamond traces shrink from a mean of 3,703 reasoning tokens at one loop pass to 1,661 at five passes, so latency, output-token billing and accuracy all move together with the configured loop/effort setting. GOVERNANCE NOTE All pre-existing Benchmark Heaven ground-truth figures were carried forward unmodified. The additional entries appended here are only those explicitly published for the exact variant Inkling 2 Mini (thinkingmachines/inkling-2-mini): GPQA Diamond 86.6%, ARC-AGI-2 49.3% and ARC-AGI-1 84.5% per ARC Prize verification, DeepSearchQA 85.3% (with compaction), MCP Atlas Public 82.7% and SWE-bench Pro v2 (Hard) 77.9%. No sibling (Inkling, Inkling 2 Small) or predecessor scores were substituted, and no value was estimated or interpolated.
26.4B total parameters; 3.4B active parameters per token at base routing, ~5.4B effective weight activations per token once the loop passes are counted. 24 stored decoder layers in four blocks of six; the two middle six-layer blocks (12 layers) are re-executed three times with shared weights while the outer blocks run once, producing 48 layer applications per token. Each MoE layer routes each token to 8 of 128 experts. Attention is a hybrid of local and global layers (GQA/MLA is not confirmed by vendor documentation). Vision Transformer encoder with image side lengths constrained to 40px–4096px. Context window 1,048,576 tokens. Reasoning is expressed through an effort/loop-depth setting rather than a documented separate reasoning-token budget; reported GPQA Diamond traces range from a mean of 3,703 reasoning tokens at one loop pass down to 1,661 at five passes. Released 20 November 2026 as open weights under Apache 2.0 with a public model card; tokenizer, data mixture, data cutoff and post-training recipe are undisclosed.
Recommended Workloads & Primary Use Cases
- •Terminal and OS-execution capability is essentially absent: Terminal-Bench 4.0 scores 5.6% and Terminal-Bench Science 1.4%, even though SWE-bench Verified reaches 79.2%. Do not route shell automation, containerised installs, or long-horizon CLI agents to this model despite its strong SWE-bench headline.
- •The 1M-token context window is a capacity ceiling, not a validated accuracy envelope. No MRCR v2, RULER or comparable long-context evaluation has been published for this exact variant, and the model's headline BrowseComp/DeepSearchQA figures are explicitly measured with harness-level compaction — accuracy on uncompacted 100k+ token tool histories and document stacks will be materially lower.
- •Reasoning effort is coupled to loop depth and is not a free parameter. Increasing loop passes compresses reasoning traces (reported GPQA Diamond mean drops from 3,703 to 1,661 tokens) while raising latency and compute per token; keep per-request max-token budgets and effort settings pinned per workload or cost and TTFT will drift unpredictably across a fleet.
- •Factual recall is unreliable and the model is over-confident in the tail: AA-Omniscience Index is −8.6 (negative, meaning more incorrect than correct answers on unknown-knowledge items) while AA-Omniscience Non-Hallucination is 63.4%. Ground every factual or numeric claim in retrieved context and validate outputs with a schema/citation check rather than trusting parametric knowledge.
- •Multimodal input has hard formatting constraints (image side 40px–4096px) and OpenRouter currently exposes the variant as text-only, so image capability must be verified per-serving-path; unsupported or oversized images will silently degrade to text-only behaviour and token accounting for images is not documented.
- •Prompt and harness sensitivity is high for agentic benchmarks: MCP Atlas is reported specifically on the Public split (82.7%) and multi-turn tool schemas, parallel call behaviour and structured-output guarantees are not documented. Validate tool-call formatting and ID round-tripping on your own harness before production rollout.
Calls to Inkling 2 Mini 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 2 Small is Thinking Machines Lab's open-weight, image-capable sparse MoE reasoning model (594B total / 17B active) with a 1,048,576-token hybrid-attention context, variable reasoning effort, and strong long-context retrieval, repo-scale coding, search-grounded research and GUI/vision grounding — at $0.30/$1.00 per 1M tokens and ~98 tps.
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.