Recursive released Flint on October 21, 2026, with downloadable weights for a coding-agent model designed to run on controlled infrastructure. The release creates a practical alternative for teams evaluating local repository work, while leaving hardware requirements, serving support, and code review as separate deployment decisions.
What actually shipped—and what API access means
Recursive’s announcement describes Flint as a mixture-of-experts model with roughly 40 billion total parameters and 2.5 billion active per token. The company released the weights under Apache 2.0 and introduced rec, its optimized local inference engine. OpenHands separately announced October 21 support after testing Flint before launch.
Availability needs three separate labels. Downloadable weights enable self-hosting; the RecursiveSI/flint repository contains a BF16 checkpoint, tokenizer, configuration and SGLang serving kit. Hosted access is a different route: OpenHands’ launch post identifies OpenRouter, and Recursive’s release documentation still directs users there or to OpenCode. Direct Recursive API access is limited to select partners, according to that documentation. A public model release therefore does not establish an unrestricted first-party API launch.
Nor should benchmarked hardware be confused with a shipped runtime feature. As checked on November 1, Recursive’s rec documentation says NVIDIA RTX Spark support for rec is still coming soon. Its published setup instructions apply to supported Macs.
The laptop release is not the 80 GB server checkpoint
Recursive’s model card puts the BF16 download at approximately 80 GB and documents one B300 or H200 GPU per replica using its SGLang kit. It also says Transformers modeling code is not yet included. The repository’s automatically generated library snippets should not override those explicit implementation notes.
The separate rec setup requires an M4 Pro, M5 Pro or M5 Max Mac, at least 48 GB unified memory, and about 30 GB of free disk space. Its installer checks hardware compatibility before downloading. The practical consequence: “local” does not mean any existing developer laptop, and the server checkpoint is not the same deployment package as the managed Mac download.
Context claims also need configuration attached. The launch post advertises support up to 512k tokens but recommends staying within 256k. The BF16 card documents 262,144 tokens shared between prompt and output. Recursive’s usage documentation further warns that local context depends on memory, quantization and cache settings. Buyers should use the limit of their actual runtime, not the largest announcement number.
Why the speed and automated-research claims need boundaries
Recursive reports more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 on a 64 GB RTX Spark at short contexts using its optimized engine. These are vendor measurements, not Nerova tests. OpenHands explicitly attributes the results to Recursive and notes that runtime, quantization, hardware and context length affect performance.
Recursive also says its Eureka Machine automated research system contributed across the model-building stack, helping a small team build Flint from scratch in under four months. That is a company account of development, not proof that autonomous, unrestricted recursive self-improvement has been achieved.
Our operational reading is narrower: faster generation can make repeated coding attempts more practical, but it does not establish lower cost per accepted change. Repository exploration, tool execution, test failures and human corrections can outweigh the time saved generating tokens.
What an engineering team should evaluate next
OpenHands describes Flint as the inference backend while its framework supplies tools and an execution environment. Recursive’s own documentation makes the same boundary explicit: the model requests actions; the harness executes them. Local inference alone therefore does not prove that the entire coding workflow stays offline.
Our suggested pilot is a bounded maintenance queue: one bug fix, a focused regression test and a small feature in a repository the team knows. Record accepted changes, reviewer minutes, retries and elapsed time. Inspect network access and tool permissions separately from model hosting, and require approval before merges or consequential commands.
This is particularly important because Recursive acknowledges that Flint may need more guidance than larger closed models. Its model card warns about faulty code, inappropriate tool calls and reasoning that exhausts output budgets. Flint’s October release creates a concrete evaluation opportunity—not evidence that teams can remove review, infrastructure planning or execution controls.