Inkling-Small: The Quarter-Size Model That Beats Its Own Flagship, and Still Trails the Index

Thinking Machines Lab spent the two weeks after shipping its first open-weights model building one about a quarter the size, and on the lab's own scorecard the smaller one wins. Inkling-Small reports 80.2% on SWE-bench Verified against Inkling's 77.6%, 89.5% on GPQA Diamond against 87.2%, 64.7% on Terminal-Bench 2.1 against 63.8%, and 31.6% on Humanity's Last Exam against 29.7% — from 276B total parameters with 12B active, rather than 975B with 41B. Of the headline numbers the two model cards share, the 975B flagship holds exactly one: AIME 2026, by 1.6 points.

Then Artificial Analysis ran both independently and put Inkling-Small at 40 on its Intelligence Index against Inkling's 41 — the smaller model a point behind. Both of those results are real, and the space between them is the most useful thing in this launch. Everything below separates what Thinking Machines reports about its own model from what someone else measured: the vendor figures come from the announcement and the two Hugging Face model cards, the independent figures from Artificial Analysis's own runs, and the pricing and context numbers from OpenRouter's live endpoint data, checked the day this was written. Where those three disagree — and on context length they do — we say so rather than pick the flattering one.

What actually shipped on July 30

Inkling-Small is a sparse mixture-of-experts transformer: 276B parameters total, 12B active per token, 42 layers, with each token routed to 6 of 256 experts plus 2 shared experts that fire on everything. Attention alternates local and global layers. Weights are on Hugging Face under Apache 2.0 with no commercial restrictions, it fine-tunes on Thinking Machines' Tinker API, and you can talk to it in text, image and audio in the Tinker Playground. The lab says it was trained on NVIDIA GB300 NVL72 systems.

Multimodality is native rather than bolted on, and the card is unusually specific about how. There is no separate vision or audio encoder: audio arrives as dMel spectrograms, images as 40×40-pixel patches pushed through a four-layer hMLP, and both get embedded straight into the same token stream as text. Input is text, images and audio; output is text only, which is worth stating plainly because "multimodal" gets read as "it can speak" often enough to cause a bad afternoon. Thinking effort is a dial with five settings — minimal, low, medium, high, xhigh — so the same weights can be run cheap or run hard.

The interesting part of the recipe is the post-training. Inkling-Small was distilled on-policy with the full Inkling as its teacher, and then given roughly two more weeks of reinforcement learning scaled up on agentic coding. That ordering explains the shape of the results: the student inherited its teacher's general competence, then got extra training on precisely the axis where it now beats the teacher.

The scoreboard Thinking Machines published

Vendor benchmarks are marketing until someone reproduces them, so read this section as what the lab claims, not as what has been verified. It is still a broad set, and it is broad in an unflattering-to-the-flagship direction.

Beyond the four shared numbers, the card fills in the rest of the picture — all Thinking Machines' own runs:

• Agentic work — 1269 Elo on GDPval-AA v2, a benchmark of realistic economic tasks rather than puzzles.

• Charts and documents — 77.4% on CharXiv RQ, rising to 81.3% when the model is allowed to write and run Python. That 3.9-point jump is a useful hint that tool access buys more on visual-reasoning work than prompt engineering does.

• Vision — 74.0% on MMMU Pro, a hair above Inkling's 73.5%.

• Audio — 90.1% VoiceBench and 77.0% MMAU, both fractionally below the flagship's 91.4% and 77.2%. Audio is one of the two places where shrinking clearly cost something.

• Safety — 98.4% StrongREJECT and 96.9% on FORTRESS benign prompts, but 71.6% on FORTRESS adversarial against the flagship's 78.0%. That is the other cost: a 6.4-point regression in adversarial robustness, which matters more than any coding gain if the model is going to sit behind a public text box.

• Forecasting — 61.3 ± 0.46 Brier Index on ForecastBench with no search access, and 0.1238 ± 0.0086 Brier score on Prophet Arena. Reporting error bars at all is more discipline than most launch posts manage.

One gap: the flagship's card lists 54.3% on SWE-bench Pro, the harder sibling of SWE-bench Verified. Inkling-Small's card does not report SWE-bench Pro. Given how loudly the Verified number is featured, its absence is the number we would most like to see filled in.

What the independent index says instead

Artificial Analysis puts Inkling-Small at 40 on version 4.1 of its Intelligence Index, one point under Inkling at 41. The index is a composite of nine evaluations — GDPval-AA v2, τ³-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience and AA-LCR — and that list explains most of the apparent contradiction. Only two of the nine overlap with the reasoning showcase on Thinking Machines' card. The rest is weighted toward agentic tool use, long-context retrieval and hallucination resistance, and a model can win the benchmarks a lab chooses to print while losing the ones a third party chooses to run. Neither party is lying; they are measuring different things.

There is also a detail in the fine print worth more than the one-point headline: Artificial Analysis lists the flagship's entry as Inkling (xhigh) — its top thinking-effort setting — while the Inkling-Small row carries no effort suffix. So the flagship's single-point lead was recorded with its reasoning dial turned all the way up. If effort-matching moved that comparison even slightly, the last argument for the larger model on capability alone would go with it.

Where the independent data is not close at all is speed and cost, and here it favours the small model by margins no benchmark table conveys:

• Output speed — 133 tokens per second versus 80 for Inkling (xhigh). Artificial Analysis ranks Inkling-Small #7 of 101 models in its class on speed, against a class median of 66.

• Time to first token — 1.63s versus 1.84s. A real but minor edge.

• Blended price per 1M tokens — $0.22 versus $0.72 on their weighting. Roughly a third the cost, for one index point less.

• Intelligence rank — #15 of 101, against a class median score of 25. Comfortably above average, nowhere near the frontier.

• Verbosity — and here is the catch. It burned 130M output tokens getting through the index, against a 100M median. Artificial Analysis's own summary calls it "notably fast, however somewhat verbose." It thinks about 30% more than typical, which quietly eats part of the per-token discount.

A small bookkeeping note, since anyone comparing sources will hit it: Artificial Analysis lists the parameter count as 266B total, while the announcement, both model cards and OpenRouter all say 276B. We have used 276B throughout. It does not change any conclusion, but it is the kind of discrepancy worth knowing about before you quote a number in a design doc.

What you can actually rent today, which is not what the spec sheet says

The gap between an open-weights announcement and a usable endpoint is where most launch coverage stops paying attention. Thinking Machines advertises a context window of up to 1M tokens, and Artificial Analysis lists 1M as well. On OpenRouter, both providers currently serving Inkling-Small cap it at 524,288 tokens — half the advertised figure. That is not a contradiction so much as a difference between what the architecture supports and what anyone has stood up in production. For contrast, the flagship Inkling does have one endpoint at the full 1,048,576 tokens, though that same endpoint limits completions to 32,768 tokens.

The rest of the live picture, read off OpenRouter's endpoint data:

• Two providers, two prices — $0.45 per 1M input and $1.20 per 1M output from one, $0.50 and $1.20 from the other. Artificial Analysis lists Thinking Machines' own first-party rate lower still at $0.30 and $1.20, with cached input at $0.06. Third-party hosting is charging a 50–67% premium on input for the same weights.

• Prompt caching — $0.10 per 1M cached input tokens via OpenRouter, against $0.17 for the flagship.

• The numerics you rent are not the numerics that were benchmarked — the model card lists BF16 and NVFP4. The cheaper endpoint serves fp8; the other does not declare a quantization at all. Vendor benchmark scores were not measured on an fp8 serving of these weights, and quantization effects on long agentic runs are exactly the kind of thing a 0.9-point Terminal-Bench margin cannot survive. If you are reproducing a number, note which endpoint you used.

• Feature coverage is uneven per provider — both endpoints expose reasoning and reasoning_effort, so the five-setting thinking dial works. But only one advertises tool calling and logprobs, and neither currently advertises response_format, the structured-output/JSON-mode parameter. The flagship Inkling has an endpoint that does. If your pipeline depends on schema-enforced JSON, this is the detail that will bite on day one, and it is a provider limitation rather than a model one.

• Completion ceiling — 262,144 tokens on the fp8 endpoint; the other declares no cap.

The cost case, on measured token counts

Per-million prices are a bad way to compare reasoning models, because the whole variable is how many tokens they burn thinking. Artificial Analysis publishes the actual spend, which is a much better instrument: getting Inkling-Small through the full Intelligence Index consumed roughly 130M output tokens and cost $192.37, working out to about $0.07 per task on their weighted average.

Set that against the same measurement for models people actually deploy: DeepSeek V4 Flash at $0.03 per task, gpt-oss-120b at $0.08, Gemini 3.6 Flash at $0.56, GLM-5.2 at $0.57, Kimi K3 at $0.86, GPT-5.6 Sol at $1.23, Claude Opus 5 at $2.34, Claude Fable 5 at $3.15. Inkling-Small is the second-cheapest model in that group and about 18× cheaper per task than GPT-5.6 Sol, which scores 59 to its 40.

There is also a cleaner way to see why the price fell where it did. Active parameters dropped from 41B to 12B, a factor of 3.4. Output price dropped from $4.05 to $1.20 per million, a factor of 3.375. The rented price of a sparse MoE is essentially just its per-token compute, and Thinking Machines charged accordingly rather than pricing to the benchmark.

One practical note on running that comparison yourself: Inkling-Small is not in the OrcaRouter catalog today, so you would rent it from its own endpoints. The closed models you would be measuring it against — GPT-5.6 Sol, Claude Opus 5, Gemini 3.6 Flash and the rest of that list — are on OrcaRouter behind a single key at 0% markup, meaning you pay each provider's list price with nothing added on top. That matters for a cost model specifically: the number you put in the spreadsheet is the number that gets billed, and when a provider cuts prices it lands on our side the same day rather than after a renegotiation. Automatic failover across providers covers the other half of an evaluation, which is the baseline arm going down mid-run and quietly poisoning your numbers.

Where it actually sits among open-weight models

Read against the flagship, Inkling-Small looks like a triumph. Read against the field, it looks like a solid mid-table open-weights model, and the launch coverage has mostly skipped the second reading.

On the Artificial Analysis Intelligence Index, the models ahead of both Inklings include Claude Opus 5 at 61, Claude Fable 5 at 60, GPT-5.6 Sol at 59, Kimi K3 at 57, Grok 4.5 at 54, GLM-5.2 and Muse Spark 1.1 at 51, and Gemini 3.6 Flash at 50. That is expected — those are frontier systems, most of them closed, several of them enormous.

The one that should actually change a decision is DeepSeek V4 Flash, which scores 50 to Inkling-Small's 40 at 284B total and 13B active — practically the same size class and the same open-weights bargain, at less than half the cost per task. If what you want is the cheapest capable open-weight model, the independent numbers point there, not here. It is also, unlike Inkling-Small, available through OrcaRouter, so testing that alternative against your own workload is a model-string change rather than a procurement exercise.

What Inkling-Small has that DeepSeek V4 Flash does not is native audio and image input in the same weights, a five-level thinking dial, and Tinker fine-tuning from the lab that trained it. Those are real differentiators for a multimodal agent, and they are the honest reason to pick it — not the index score.

Who should move, and who should stay put

The decision splits cleanly, which is unusual enough to be worth stating as such.

• Move from Inkling to Inkling-Small if you are running coding agents. The vendor numbers favour the small model on SWE-bench Verified and Terminal-Bench, the extra agentic RL is the documented reason, and it costs about a third as much and returns tokens 1.66× faster. Paying 3.3× for one index point measured at maximum thinking effort is a hard case to make.

• Move if cost per reasoning task is the binding constraint and you can live with 40 on the index. $0.07 per task with an $0.06-per-million cache-hit rate on the first-party endpoint is aggressive pricing for a model with native audio and vision.

• Move if you are fine-tuning. A 276B/12B model is dramatically cheaper to adapt and serve than 975B/41B, and Tinker supports both.

• Stay on Inkling if you need long audio. This is the sharpest regression in the release and it is buried in the model cards: the flagship recommends audio inputs under 20 minutes, while Inkling-Small's card recommends under 2 minutes. A tenfold cut. If you were transcribing or reasoning over meetings, the small model is not a drop-in replacement at any price.

• Stay if you need context beyond 512K from a hosted endpoint, or completions longer than 262K tokens. Today only the flagship has an endpoint serving the full million.

• Stay if you need schema-enforced JSON without writing your own validation-and-retry loop, until a provider ships response_format for the small model.

• Stay if adversarial robustness is load-bearing. 71.6% versus 78.0% on FORTRESS adversarial is the wrong direction for anything user-facing, and it is the lab's own number.

Three questions the launch coverage left open

Is Inkling-Small just a distilled Inkling?

Partly, and the part that isn't is the part that matters. On-policy distillation with Inkling as teacher was one stage of post-training, so a lot of the general capability really is inherited. But this is a separately architected model — 42 layers against the flagship's 66, with the same routing topology of 6 active experts out of 256 plus 2 shared, which means the experts themselves are narrower rather than fewer. And it received about two weeks of agentic-coding RL that the teacher never got. That last stage is why a distilled student outscores its teacher on coding benchmarks, which otherwise shouldn't happen.

Can I realistically self-host it?

Only on serious hardware. 276B parameters is roughly 276GB of weights at 8-bit and about 552GB at BF16, before any KV cache — a multi-GPU node either way, not a workstation. The upside of sparse MoE is inference cost, not memory footprint; you store all 276B and compute with 12B. The card names SGLang, vLLM, TokenSpeed, Unsloth and Hugging Face Transformers as supported serving paths and lists BF16 and NVFP4 numerics. Note also that both hosted endpoints today serve a quantized version, so "the same model" self-hosted at BF16 will not behave identically to the API you tested against.

Does the 975B Inkling still have a reason to exist?

Three, and they are all narrow: the full 1M-token served context, audio inputs longer than two minutes, and better adversarial safety behaviour. Notably, "it is smarter" is not confidently on that list — one index point at maximum thinking effort, against a set of vendor benchmarks the smaller model mostly wins, is not a capability argument. Two weeks after launching a 975B flagship, Thinking Machines released the model that makes it hard to recommend. That is either unusual candour or unusual speed, and either way it is a good sign about the lab.

What to watch next

Three things will settle how this release ages. Whether an endpoint appears serving the advertised 1M context, which would remove the flagship's clearest remaining advantage. Whether anyone publishes an effort-matched independent comparison of the two models, which is the only way to know if that one index point is real. And whether the 2-minute audio recommendation is a training limitation or a serving default — because if it is the latter, the single biggest reason to stay on the larger model evaporates. Until then, the honest summary is that Thinking Machines shipped a model that is a third the price, two-thirds faster, better at coding on its own evidence, marginally behind on independent aggregate scoring, and clearly behind a same-size rival that most of the coverage did not mention.