Seventeen billion parameters show up to work. Five hundred ninety-four billion are on the payroll. That is the bargain Thinking Machines Lab is selling with Inkling 2 Small, and as of November 5, 2026 it is a hiring decision, not a spec-sheet flex.
On Terminal-Bench 4.0 the model scores 13.1%. DeepSeek-V4.1-Flash-1017 scores 33.8% on the same test. GLM-5.3-Flash scores 32.8%. If the job is an AI agent that lives in a shell, the payroll is the wrong résumé.
Inkling 2 Small is the first model of a second Inkling generation, not a quiet patch on the last checkpoint. It takes text and images and returns text. There is no audio at launch. The context window is listed at 1 million tokens. Artificial Analysis notes that a maximum context figure covers input and output together, and that output limits are often lower, so do not budget a million tokens of answer until an output cap is published. Reasoning effort is adjustable. The lab’s published task scores were run at effort 0.99, labeled “xhigh” (the dial nearly all the way up), and at temperature 1.0. What you give up by turning that dial down is not documented.
The numbers that should decide a pilot
Figures drawn from Artificial Analysis on November 5, 2026 put Inkling 2 Small at 31 on the Intelligence Index, up from 13 for Inkling Small. On that same read, DeepSeek-V4.1-Flash-1017 scores 33 and GLM-5.3-Flash scores 28. The jump against its own predecessor is the striking part. Against current peers it is a point or two, not a lead change. A composite blends terminal work, coding, knowledge tasks, and more. It will flatter a model that is excellent at three things and poor at one, which is roughly the shape here.
The sharpest lab number is on hard software repair, and it needs one caveat, not five. Thinking Machines reports 86.3% on SWE-Bench Pro v2’s Hard split at that xhigh setting. Scale, which maintains the benchmark, describes the Hard split as 51 tasks drawn from verified failure modes: multi-file dependencies, subtle system bugs, long-horizon reasoning. The 86.3% figure is lab-reported pass@1 (one attempt, not the best of several) and second only to Kimi K3 inside the lab’s own table. Those Hard scores do not appear in Scale’s write-up of the benchmark. It is not a private-set result. Scale treats a private set of 272 tasks as the cleaner measurement, because the public set of 642 has lived on the open web, and cites a 17.8-point public-to-private gap for Claude Opus 5 as the example. A strong public score can mix skill with training-time recall.
It does not sweep the coding tests either. On DeepSWE v1.1 the lab comparison shows 72.1%, behind DeepSeek-V4.1-Flash-1017 at 77.0%. A September DeepSeek checkpoint is listed separately at 74.2%. GLM-5.3-Flash sits at 63.4% in that table.
One naming trap, then the footnotes can close. Scale’s harness notes discuss a model it calls “Inkling” and groups with open-weight systems. Inkling 2 Small launched closed. A shared family name, and a shared “xhigh” label, are not identification. Do not paste those resolved-task counts onto this release.
The miss that should slow a pilot is specific. Terminal-Bench 4.0 is the 13.1% result, against 33.8% for DeepSeek-V4.1-Flash-1017 and 32.8% for GLM-5.3-Flash. Those numbers come from two different peers. A different test, Terminal-Bench 2.1, shows 85.1% for Inkling under a Pi harness in the lab column, with a September DeepSeek checkpoint at 90.6% and GLM-5.3-Flash at 84.3%. The published materials do not say what version 4.0 adds or hardens relative to 2.1: task mix, harness, or contamination control. They are not the same exam. If someone quotes you a terminal number, ask which version before the pilot gets a budget. Artificial Analysis includes Terminal-Bench 4.0 in the Intelligence Index. It is the 13.1% row that sits inside the composite of 31, not the 85.1%.
The invoice is not the sticker price
Token prices on that November 5 Artificial Analysis read are $0.30 per million input tokens and $1.00 per million output tokens. Output is cheaper than Inkling Small, listed at $1.20. DeepSeek-V4.1-Flash-1017 matches the $0.30 input price and charges $1.20 on output. GLM-5.3-Flash is the cheap seat, at $0.15 and $0.50.
Sticker price is not the invoice. Artificial Analysis builds cost per task from a weighted blend of input, cache-hit, cache-write, reasoning, and answer tokens. On that measure Inkling 2 Small runs $1.18 per index task, against $0.60 for DeepSeek and $0.58 for GLM. Roughly double, for a composite that does not lead the trio.
Time per task is output tokens divided by end-to-end output speed. That is model-output time. It may not include tool waits, sandbox time, or a human reading the diff, so 21.5 minutes is not necessarily wall-clock. DeepSeek finishes the same style of index task in 13.3 minutes on that clock. GLM takes 37.7. Slower than one peer, faster than the other. “Slow” depends on who is standing next to it.
All three sit in a similar active-parameter neighborhood, as cited from their model cards in that comparison: about 17 billion here, 16 billion on decode for DeepSeek with 8 billion on prefill, and 18 billion for GLM. Totals diverge more: 594 billion, 552 billion, and 320 billion. All three advertise a 1 million-