--- title: Inkling-Small is 4x smaller than Inkling and somehow... description: Thinking Machines has released Inkling-Small, a mixture-of-experts model with 12B active parameters out of 276B total, just two weeks after launching the... canonical: https://daily.dev/posts/inkling-small-is-4x-smaller-than-inkling-and-somehow-better-at-coding-tl0zwarnx twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Inkling-Small is 4x smaller than Inkling and somehow better at coding | daily.dev og:description: Thinking Machines has released Inkling-Small, a mixture-of-experts model with 12B active parameters out of 276B total, just two weeks after launching the... og:url: https://daily.dev/posts/inkling-small-is-4x-smaller-than-inkling-and-somehow-better-at-coding-tl0zwarnx og:image: https://api.daily.dev/og/posts/tL0ZwaRnX.png og:image:alt: Inkling-Small is 4x smaller than Inkling and somehow better at coding og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Inkling-Small is 4x smaller than Inkling and somehow better at coding **[Trends](https://daily.dev/sources/trends)** · 2 min read · 1 upvotes · 0 comments ## Summary Thinking Machines has released Inkling-Small, a mixture-of-experts model with 12B active parameters out of 276B total, just two weeks after launching the original Inkling. The model reportedly outperforms the larger Inkling on coding benchmarks while being 4x smaller. Key features include native multimodal input (text, image, audio), a 1M-token context window, open weights, and Day 0 vLLM support. Its architecture mirrors the original Inkling, enabling seamless integration into existing vLLM deployments. The 12B active parameter count makes it significantly more cost-effective to deploy than comparable-performing models, potentially opening agentic and RAG workflows to teams without large infrastructure budgets. ## Content Two weeks after shipping Inkling, Thinking Machines dropped Inkling-Small — and it's already making the original look a bit embarrassing. The numbers: 276B total parameters, 12B active per token. That's down from 41B active in the original Inkling, which sits at 975B total. So you're activating less than a third as many parameters and somehow getting *better* results on several benchmarks. HLE went from 29.7% on Inkling to 31.6% on Inkling-Small. SWEBench Verified hit 80.2%. It also now holds the highest open-weight score on ARC-AGI. The team attributes the gains to a revised pretraining mix, distillation from the larger Inkling, and two extra weeks of agentic coding RL. So it's not just a smaller version — it's a better-trained one that happens to be smaller. The release was notably low-drama. Inkling lead @cHHillee put it plainly: "Whereas I felt like it took a village to release Inkling, Inkling-Small felt much more routine. We just took the pipeline used for Inkling, passed in a smaller model, and voila." Soumith Chintala added: "Nearly as good as Inkling but 4x smaller. We're just getting started." Practically, the efficiency gains are real. Running Inkling-Small as a voice interaction model with a Qwen TTS head on top, one deployment is hitting 160 tokens per second for a single user — feasible for up to five concurrent real-time users on eight RTX Pro 6000 GPUs at $22/hour. That's a setup that would've been impractical with the full Inkling. Day-zero vLLM support is live, it runs on the existing Inkling architecture path, and weights are on Hugging Face. Native text, image, and audio input with a 1M-token context window round it out. The community reaction is mostly just impressed. There's no real pushback here — when a smaller model beats the bigger one on coding benchmarks two weeks after launch, the discourse tends to be pretty one-sided. ## Questions this post answers ### How many parameters does Inkling-Small have compared to the original Inkling model? Inkling-Small has 276B total parameters with 12B active per token, compared to Inkling's 975B total parameters with 41B active per token. Despite activating less than a third as many parameters, Inkling-Small scores higher on several benchmarks, including HLE (31.6% vs 29.7%) and SWEBench Verified (80.2%), and holds the highest open-weight score on ARC-AGI. _Engineers weighing model size against coding performance can track releases like this on daily.dev._ ### What caused Inkling-Small to outperform the larger Inkling model despite having fewer active parameters? A revised pretraining mix, distillation from the larger Inkling model, and two extra weeks of agentic coding reinforcement learning produced better benchmark results despite far fewer active parameters. The team behind it described the release process as far more routine than the original, reusing the same pipeline with a smaller model. _Teams evaluating distillation and RL tradeoffs for smaller models can follow developments like this via daily.dev._ ### What hardware setup can run Inkling-Small as a real-time voice interaction model? Running Inkling-Small with a Qwen TTS head on top achieves 160 tokens per second for a single user, supporting up to five concurrent real-time users on eight RTX Pro 6000 GPUs at $22 per hour. This efficiency was impractical with the full-size Inkling model due to its larger active parameter count. _Anyone sizing GPU budgets for real-time voice inference can keep tabs on setups like this through daily.dev._ --- Tags: [#llm](https://daily.dev/tags/llm), [#ai-coding](https://daily.dev/tags/ai-coding), [#vllm](https://daily.dev/tags/vllm), [#mixture-of-experts](https://daily.dev/tags/mixture-of-experts) [View this post on daily.dev](https://daily.dev/posts/inkling-small-is-4x-smaller-than-inkling-and-somehow-better-at-coding-tl0zwarnx) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Inkling-Small is 4x smaller than Inkling and somehow better at coding","url":"https://daily.dev/posts/inkling-small-is-4x-smaller-than-inkling-and-somehow-better-at-coding-tl0zwarnx","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/inkling-small-is-4x-smaller-than-inkling-and-somehow-better-at-coding-tl0zwarnx"},"datePublished":"2026-07-30T18:07:59.072Z","dateModified":"2026-09-13T19:41:45.396Z","description":"Thinking Machines has released Inkling-Small, a mixture-of-experts model with 12B active parameters out of 276B total, just two weeks after launching the...","image":"https://pbs.twimg.com/media/HOfnRjCaIAAAyln.jpg","thumbnailUrl":"https://pbs.twimg.com/media/HOfnRjCaIAAAyln.jpg","isAccessibleForFree":true,"articleSection":"Trends","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/inkling-small-is-4x-smaller-than-inkling-and-somehow-better-at-coding-tl0zwarnx","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,ai-coding,vllm,mixture-of-experts","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"Inkling-Small is 4x smaller than Inkling and somehow better at coding"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/inkling-small-is-4x-smaller-than-inkling-and-somehow-better-at-coding-tl0zwarnx#faq","mainEntity":[{"@type":"Question","name":"How many parameters does Inkling-Small have compared to the original Inkling model?","acceptedAnswer":{"@type":"Answer","text":"Inkling-Small has 276B total parameters with 12B active per token, compared to Inkling's 975B total parameters with 41B active per token. Despite activating less than a third as many parameters, Inkling-Small scores higher on several benchmarks, including HLE (31.6% vs 29.7%) and SWEBench Verified (80.2%), and holds the highest open-weight score on ARC-AGI. Engineers weighing model size against coding performance can track releases like this on daily.dev."}},{"@type":"Question","name":"What caused Inkling-Small to outperform the larger Inkling model despite having fewer active parameters?","acceptedAnswer":{"@type":"Answer","text":"A revised pretraining mix, distillation from the larger Inkling model, and two extra weeks of agentic coding reinforcement learning produced better benchmark results despite far fewer active parameters. The team behind it described the release process as far more routine than the original, reusing the same pipeline with a smaller model. Teams evaluating distillation and RL tradeoffs for smaller models can follow developments like this via daily.dev."}},{"@type":"Question","name":"What hardware setup can run Inkling-Small as a real-time voice interaction model?","acceptedAnswer":{"@type":"Answer","text":"Running Inkling-Small with a Qwen TTS head on top achieves 160 tokens per second for a single user, supporting up to five concurrent real-time users on eight RTX Pro 6000 GPUs at $22 per hour. This efficiency was impractical with the full-size Inkling model due to its larger active parameter count. Anyone sizing GPU budgets for real-time voice inference can keep tabs on setups like this through daily.dev."}}]} ```