--- title: Recursive SI introduces Flint, a fast local coding model description: Richard Socher announced Flint, the first model from Recursive SI. He describes it as a fast coding model that runs locally. The model was built using the... canonical: https://daily.dev/posts/recursive-si-introduces-flint-a-fast-local-coding-model-htkpkjuut twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Recursive SI introduces Flint, a fast local coding model | daily.dev og:description: Richard Socher announced Flint, the first model from Recursive SI. He describes it as a fast coding model that runs locally. The model was built using the... og:url: https://daily.dev/posts/recursive-si-introduces-flint-a-fast-local-coding-model-htkpkjuut og:image: https://api.daily.dev/og/posts/htkPkJuuT.png og:image:alt: Recursive SI introduces Flint, a fast local coding model 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. # Recursive SI introduces Flint, a fast local coding model **[Collections](https://daily.dev/sources/collections)** · 4 min read · 0 upvotes · 0 comments ## Summary Richard Socher announced Flint, the first model from Recursive SI. He describes it as a fast coding model that runs locally. The model was built using the company's automated research system. The announcement, reshared by @srush_nlp and @_akhaliq, gives no benchmarks, model size, or release date, and doesn't explain how the automated research system works. ## Content Recursive Superintelligence has released Flint, a free coding model small enough to run on a laptop. The weights are on Hugging Face under Apache 2.0, with GGUF, bf16, int4 and NVFP4 versions. The company says it built Flint from scratch in under four months, largely with its own automated research system, the Eureka Machine. ## What it is Flint is a 40B mixture-of-experts model with only 2.5B parameters active per token. That is why it's fast on consumer hardware. It runs through llama.cpp, LM Studio, Unsloth, vLLM, SGLang, or Recursive's own inference engine. Unsloth has published GGUFs and says Flint works well for tool calls and agentic coding in Unsloth Desktop. The Eureka Machine is a loop that proposes ideas, implements them, tests them, and uses the results to pick what to try next. Recursive says it contributed to nearly every part of the stack: data, training recipes, and inference kernels. Specific results credited to it: - A sparser architecture. - Low-bit training. Compressing the output layer to 4 bits shrank the model by 71% with almost no quality loss, according to Richard Socher. - Custom GPU kernels. On an RTX Spark, one set sped Flint up by 55% without changing the weights. ## Speed Recursive reports more than 125 tokens per second on an M4 Pro MacBook Pro at short contexts, over 150 on Nvidia's RTX Spark laptops, and more than 220 on a 64GB RTX Spark. TechCrunch reported that similar models such as Google's Gemma 4 managed 58 to 69 on the same Mac in the company's tests. One comparison puts a dense Qwen3-27B at about 13 tokens per second on a MacBook versus about 130 for Flint. Long contexts hurt. One reviewer saw speed fall to about 9 tokens per second at 115,000 tokens. Recursive's own chart on a 48GB MacBook claims roughly 7x the speed of the strongest local alternative over a 128K-token session, at about 81% average quality against about 85%. ## Quality In Recursive's tests, Flint scored 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1. The company says it tops MoE models under 50B on three major agentic coding benchmarks. It still trails the dense Qwen3-27B on some of them: Terminal-Bench is about 55 versus 73, while LiveCodeBench is close. Hands-on reports are mixed. One reviewer, testing in Pi and OpenCode through the Panda Code app, had it fix a six-bug shopping cart library correctly in 7.5 minutes and build a working link shortener API with 14 passing tests in 9 minutes. A frontend landing page it produced had a CSS sizing bug. Another tester found it answered a counting question with 12, 14, then 17, when the correct answer was 43, and it never checked its work. Fast is not the same as careful. ## The company Recursive raised $650 million at a $4.65 billion valuation and has a team of roughly 30 people. It shipped the model, the inference engine, and the research system that helped build both. ## Why people are paying attention Flint arrived the same day as Mirendil-Base, a 100B-500B coding model that Mirendil says it pretrained in 3.5 months with one researcher and under $10M of compute. Agents wrote over 99% of the code and ran the experiments, while the researcher chose directions. Researcher Robert Lange noted that the framings differ. Mirendil cuts the people, compute and time per model with a human steering agents. Recursive wants to automate more of the research loop itself. Both start with coding models, on the bet that better coding speeds up AI research. I'd treat "first recursive self-improvement model" with some caution. The benchmarks are the company's own, and the counting failure is a reminder of the gap with the best dense models. Still, a coding model that runs at this speed on a laptop with no API bill is something local users haven't had before, and it's the first time I've seen an automated research loop produce a model I can actually download. ## Questions this post answers ### What is Flint, the new open-weight coding model from Recursive SI? Flint is a 40B parameter mixture-of-experts coding model with 2.5B active parameters, released by Recursive SI under the Apache 2.0 license on Hugging Face. It runs at over 125 tokens per second on an M4 Pro MacBook Pro at short contexts and tops its class among MoE models under 50B parameters on three agentic coding benchmarks. _Developers weighing local versus hosted coding models can track releases like Flint on daily.dev._ ### How did Recursive SI's Eureka Machine research system help build Flint? The Eureka Machine, an automated research system, contributed to nearly every part of the stack: data, training recipes, and inference kernels. It found a sparser architecture and a low-bit training approach that shrank the model by 71% with almost no quality loss, plus custom inference kernels, work that took the team from scratch to release in under four months. _Anyone following automated AI research tooling can keep up with cases like this on daily.dev._ ### How fast is Flint's inference on consumer hardware compared to cloud-hosted coding models? Flint runs locally at over 125 tokens per second on an M4 Pro MacBook Pro at short contexts, running entirely on a laptop rather than a billed cloud server. It is a 40B parameter MoE model with 2.5B active parameters, positioned as a fast local alternative to token-billed coding models. _Developers deciding between local and cloud inference for coding assistants can follow benchmarks like this on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 3 discussions and 17 comments across x (as of 2026-10-27). **TL;DR:** Early reactions focus on Flint's local inference speed and quantization options, with people impressed by the decode speed and crediting the custom kernel work for the reported speedup, while others probe practical details like memory footprint and tool-call reliability at lower bit widths. **Sentiment:** 55% positive · 35% mixed · 10% skeptical **The case for** - The 55% speedup is seen as coming from kernel engineering rather than retraining, which some find the most notable detail of the launch. - Decode speed on local hardware (220 tokens/sec) is called out as impressive. - Memory headroom looks comfortable even with a full long-context session on a 64GB card. **The pushback** - Some express skepticism that the reported speed is achievable without seeing it demonstrated firsthand. - Questions remain about whether tool-call reliability degrades at lower bit-width quantizations in multi-step agent loops. **By community** - x (positive): Replies highlight the kernel-driven speedup and fast local decode speed as the standout takeaways, with practical follow-up questions about quantization and browser-agent latency. **Open questions** - Does tool-call reliability hold up at lower bit widths (e.g. on 24GB setups) compared to 8-bit in multi-step agent loops? - How does Flint perform turn-by-turn in real browser-use agent workflows under a 5-second budget? **Highlights** > @testingcatalog a 55% speedup with the weights untouched means the gain came from kernels, not training. that's the number i'd screenshot from this launch > — [fireplyai on x · 1 points](https://x.com/fireplyai/status/2112984299981217960) > @testingcatalog Memory won't be the limit there: the NVFP4 GGUF is 22.86GB and a full 262K context adds roughly 7.5GB at f16, under half of the 64GB. > — [SelfHostedSI on x](https://x.com/SelfHostedSI/status/2113049315132236036) > @testingcatalog 220 tokens per second on a local 40B moe? that's wild, i'd believe it was faster if i saw it. > — [chenzhangreal on x](https://x.com/chenzhangreal/status/2113112279256306032) > @danielhanchen If you're on a 16GB card, UD-IQ3_XXS at 14.54GB is the biggest Flint quant that still fits in VRAM. The Q3_K_XL is already 18.49GB. > — [SelfHostedSI on x](https://x.com/SelfHostedSI/status/2113077040048144532) > @danielhanchen Useful. On a 24GB machine, how does tool-call reliability hold up in the dynamic GGUF at lower bit widths? Do you see more malformed calls in multi-step agent loops than at 8-bit? > — [BongJupite43260 on x](https://x.com/BongJupite43260/status/2113101007534326091) **Source threads** - [x](https://x.com/testingcatalog/status/2112983874309628360) · 1 points · 5 comments - [x](https://x.com/danielhanchen/status/2113048469103673799) · 0 points · 12 comments - [x](https://x.com/svpino/status/2113128457823764725) · 0 points · 0 comments --- Tags: [#local-ai](https://daily.dev/tags/local-ai) [View this post on daily.dev](https://daily.dev/posts/recursive-si-introduces-flint-a-fast-local-coding-model-htkpkjuut) ```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":"Recursive SI introduces Flint, a fast local coding model","url":"https://daily.dev/posts/recursive-si-introduces-flint-a-fast-local-coding-model-htkpkjuut","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/recursive-si-introduces-flint-a-fast-local-coding-model-htkpkjuut"},"datePublished":"2026-10-21T16:35:37.209Z","dateModified":"2026-10-27T10:48:07.154Z","description":"Richard Socher announced Flint, the first model from Recursive SI. He describes it as a fast coding model that runs locally. The model was built using the...","isAccessibleForFree":true,"articleSection":"Collections","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":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/recursive-si-introduces-flint-a-fast-local-coding-model-htkpkjuut","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"local-ai","timeRequired":"PT4M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Recursive SI introduces Flint, a fast local coding model"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/recursive-si-introduces-flint-a-fast-local-coding-model-htkpkjuut#faq","mainEntity":[{"@type":"Question","name":"What is Flint, the new open-weight coding model from Recursive SI?","acceptedAnswer":{"@type":"Answer","text":"Flint is a 40B parameter mixture-of-experts coding model with 2.5B active parameters, released by Recursive SI under the Apache 2.0 license on Hugging Face. It runs at over 125 tokens per second on an M4 Pro MacBook Pro at short contexts and tops its class among MoE models under 50B parameters on three agentic coding benchmarks. Developers weighing local versus hosted coding models can track releases like Flint on daily.dev."}},{"@type":"Question","name":"How did Recursive SI's Eureka Machine research system help build Flint?","acceptedAnswer":{"@type":"Answer","text":"The Eureka Machine, an automated research system, contributed to nearly every part of the stack: data, training recipes, and inference kernels. It found a sparser architecture and a low-bit training approach that shrank the model by 71% with almost no quality loss, plus custom inference kernels, work that took the team from scratch to release in under four months. Anyone following automated AI research tooling can keep up with cases like this on daily.dev."}},{"@type":"Question","name":"How fast is Flint's inference on consumer hardware compared to cloud-hosted coding models?","acceptedAnswer":{"@type":"Answer","text":"Flint runs locally at over 125 tokens per second on an M4 Pro MacBook Pro at short contexts, running entirely on a laptop rather than a billed cloud server. It is a 40B parameter MoE model with 2.5B active parameters, positioned as a fast local alternative to token-billed coding models. Developers deciding between local and cloud inference for coding assistants can follow benchmarks like this on daily.dev."}}]} ```