Recursive's Flint Coding Model on RTX Spark: 150 vs. 220 Tokens per Second Explained

Bottom line: TechCrunch reports Flint exceeding 150 tokens per second on an RTX Spark, while NVIDIA cites more than 220 on a 64GB RTX Spark configuration. Both sources describe a 40-billion-parameter model with about 2.5 billion parameters active per token.

Two pages published on October 21 both describe Flint, and they do not describe the same speed. TechCrunch says that on Nvidia’s RTX Spark, Flint topped 150 tokens per second. NVIDIA’s blog the same day says Recursive Flint delivers over 220 tokens per second on a 64GB RTX Spark. One hundred fifty and 220. That is not a rounding error. It is either a different machine, a different test, or a number that moved between the company’s test and the chip vendor’s sentence. The pages do not say which.

The size, at least, lines up. TechCrunch calls Flint a mixture-of-experts model that uses 2.5 billion of its 40 billion parameters at a time. NVIDIA’s blog calls it a 40-billion-parameter mixture-of-experts model with 2.5 billion active parameters. WIRED, also October 21, says Flint has 40 billion parameters and that Will Knight tested it on a Mac with 48GB of RAM. A card that says “40 billion, so it needs a data-center GPU” has skipped the active count the other two pages printed, and it has skipped the fact that a journalist ran it on a 48GB Mac.

We already split a 5x that was three different sentences. Super VL’s 5x was not the Skydance 5x. We already split a parameter count from a file size. EmbeddingGemma 2 is 740 million, and the context is 8K. We already had a speed claim that was not the number in the other column. FLUX 3 Max’s 22x was 12.45 seconds, not 10.95. Flint is the laptop version of that habit. The weights are the easy part. The rate is not.

Flint’s 150 and 220 are not one RTX Spark sentence

TechCrunch’s speed section is Recursive’s tests, as TechCrunch relays them. On a MacBook Pro with Apple’s M4 Pro chip, Flint generated about 131 tokens per second. Similar models on that machine, including Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning, managed 58 to 69 tokens per second. TechCrunch calls 131 roughly twice that range. Twice 58 is 116. Twice 69 is 138. 131 sits inside that span. “Roughly twice” is a fair caption for the Mac comparison TechCrunch printed. It is not a caption for 220.

The RTX Spark line in that piece is one clause: Flint topped 150 tokens per second. It does not say 64GB. It does not say which Spark SKU. It does not say whether “topped” means a peak or a sustained rate. NVIDIA’s blog, timestamped in the search record at 10:50 GMT on October 21, specifies a 64GB RTX Spark and says over 220 tokens per second. TechCrunch’s story is timestamped 20:27 GMT the same day, so the 150 was published later, not earlier. A later story with a lower rate is not automatically a correction. It can be a different box. “RTX Spark” and “64GB RTX Spark” are not proven to be the same noun just because both contain Spark.

A buyer comparing a 48GB Mac, which is the machine WIRED names, to a 64GB Spark, which is the machine NVIDIA names, is already comparing two memory figures. TechCrunch’s M4 Pro rate, 131, is a third machine. Three rates, three boxes, unless a page says otherwise.

Nemotron 3.5 Lightning, the model TechCrunch puts in the 58-to-69 band on the M4 Pro, is not the Super VL this site wrote up on October 19. Lightning is the name in TechCrunch’s comparison sentence. Super VL was a 120-billion-parameter multimodal agent model with a 5x claim that did not survive contact with the next sentence. If a dashboard labels the Mac result “Nemotron 3.5” and you paste Super VL’s ranking next to Flint’s 131, you have mixed a laptop token rate with a data-center marketing multiple. TechCrunch’s 58 to 69 is a rate on one Mac, for models it called similar, including Lightning. It is not a leaderboard placement for the VL.

40 billion is the total. 2.5 billion is what the token uses

The active-parameter line is the one WIRED’s excerpt drops, and it is the line that decides whether “40 billion” means a laptop model or a server model.

TechCrunch: just 2.5 billion of its 40 billion settings, at a time. NVIDIA: 2.5 billion active parameters. 2.5 divided by 40 is 6.25%. That percentage is my division. Neither page prints 6.25%. They print the two counts. A spec sheet that keeps only 40 billion will overstate the slice that runs per token, if those pages are right. A spec sheet that keeps only 2.5 billion will understate the weights you download. WIRED’s test note is the practical check on the download. Knight ran it on 48GB of RAM, which he calls a healthy amount but only enough for a modest-size model like Flint. He contrasts that with the biggest models, several trillions of parameters, which he says can only run on racks of industrial servers. 48GB is not a trillion-parameter rack. It is also not a promise that every 40-billion-parameter model fits in 48GB. Flint’s active count is why TechCrunch says it is light enough for an ordinary laptop. WIRED’s 48GB test is one laptop-class machine, not every laptop.

NVIDIA’s same post lists a different active count for a different model, and the two will get welded. Liquid AI’s LFM2.5-24B-A2B activates just 2 billion of its 24 billion parameters per token. Two billion active is not 2.5 billion active. Twenty-four billion total is not 40 billion total. 2 divided by 24 is 8.33%, my division again, and a higher active fraction than Flint’s 6.25% on a smaller model. If a roundup says “the new local models activate about 2 billion parameters,” it has rounded Flint’s 2.5 down and erased LFM’s 24.

Developers can download the open weights on Hugging Face under the Apache 2.0 license and run them on RTX Spark and DGX Spark systems, and on GeForce RTX and RTX PRO GPUs. TechCrunch says Flint runs on the user’s computer rather than a company’s cloud servers, so developers can use it offline and keep code private. “Can” is doing work. Offline is a property of a local run, not a property of every wrapper someone will ship around the weights. A hosted demo of an Apache-licensed model is not offline. We have already had a local flag that was not an offline product.

SWE-Bench Verified is not SWE-Bench Pro v2

The benchmark names split, and neither page prints the score.

TechCrunch says that on SWE-Bench Verified, Flint outperformed other models under 50 billion parameters. It does not print Flint’s percent. It does not name the models it beat. “Under 50 billion” is a size fence. Flint’s 40 billion total sits under that fence. A 27-billion model also sits under it. Outperformed, with no table, is a ranking you cannot rebuild from the sentence.

The next sentence is a different test. On Recursive’s own average across six tests, TechCrunch says, only one model scored higher than Flint: Alibaba’s Qwen3.8 27B. Six tests, an average, one model ahead. That is not “best on SWE-Bench Verified.” A 27-billion model beating Flint on Recursive’s average does not contradict a claim that Flint beat other models under 50 billion on one named benchmark. It does contradict a headline that says Flint leads every open coding model. TechCrunch did not write that headline. The sentences it did write are a win on one benchmark against an unnamed under-50-billion set, and a second-place on the company’s own six-test average, behind Qwen3.8 27B.

NVIDIA’s blog says best-in-class scores in agentic coding benchmarks like SWE-Bench Pro v2, Terminal-Bench 2.1, and SWE-bench verified. Three names. No scores. “Like” is a list of examples, not a table. SWE-Bench Pro v2 is not that benchmark. Terminal-Bench 2.1 is not that benchmark. Best-in-class, with no number and no rival named, is the same shape as the 5x that did not name its runner-up.

WIRED says Recursive shared benchmark scores suggesting Flint outperforms other models of similar size and architecture on coding tasks. No table in the excerpt. “Similar size and architecture” is Socher’s comparison as Knight relays it, not TechCrunch’s under-50-billion fence, and not NVIDIA’s three benchmark names. A leaderboard screenshot that is not from these pages is a fourth document.

TechCrunch also says the company acknowledged Flint would need more guidance than the big paid coding assistants, and suggested breaking large tasks into smaller steps. That sentence is the counterweight to every best-in-class line. Knight, via TechCrunch, used Flint on the 48GB Mac to build a to-do list app and a tool that scours the web for research, and found that it nailed those requests, where smaller coding models often ran into errors or were slow compared with Claude or Codex. A to-do list is not SWE-Bench Pro v2. Both can be true. One is a journalist’s afternoon. The other is a benchmark name without a score.

The Eureka Machine is the method, not a second model

Socher, identified by WIRED as Recursive’s CEO and cofounder, told Knight that nearly every aspect of the entire model pipeline was done and improved upon by AI. “Nearly” is in the quote. A headline that drops nearly has upgraded a qualified claim. He told WIRED the system, dubbed the Eureka Machine, devised a new way of training the base model, new architectures, and code optimized for specific silicon. That line is WIRED’s interview.

The longer-term goal, Socher told WIRED, is full recursive self-improvement, with humans overseeing the process and able to stop it. That is a goal. Flint is a released coding model. A goal and a release are different tenses. TechCrunch says each new model should make the Eureka Machine a better researcher, helping it build the next one faster. “Should” is TechCrunch’s relay of the company’s claim. Nobody in these excerpts printed one.

Recursive Superintelligence is the name on NVIDIA’s blog. Recursive is the name on TechCrunch and WIRED. If the credit line on Hugging Face says something else, use that line for the credit, not my assumption.

Muse Spark Open’s terabyte is TechCrunch’s comparison, not Flint’s requirement

TechCrunch uses Meta’s Muse Spark Open as the size contrast. It says Muse Spark Open, released last week, uses 816 billion parameters and needs at least a terabyte of memory. Flint’s 40 billion, with 2.5 billion active, is the model TechCrunch says is far smaller. The 816 billion and the terabyte are TechCrunch’s sentences about a different model. They are not Flint’s spec.

The same NVIDIA post has other day-zero models that will get pulled into a Flint recap if the recap is “what ran on Spark this week.” Black Forest Labs’ FLUX 3 Image Klein is a 5.6-billion-parameter open-weights image model. NVIDIA-optimized builds, the blog says, run on GeForce RTX GPUs with as little as 8GB of VRAM and create a 1-megapixel image in about four seconds on an RTX 5090. Four seconds for a 1-megapixel Klein image is not the 12.45-second figure this site logged for FLUX 3 Max. Klein is 5.6 billion parameters and an image model. Max was a different product sentence. MiniMax M3.1 Flash, on the same blog, is described as a new open model built for DGX Station and DGX Spark.

Forty billion total. 2.5 billion active. Apache 2.0, if you are reading NVIDIA. About 131 tokens per second on an M4 Pro MacBook Pro, if you are reading TechCrunch. Over 220 on a 64GB RTX Spark, if you are reading NVIDIA. Topped 150 on “RTX Spark,” if you are reading TechCrunch. WIRED’s 48GB test does not pick a side in that rate fight, because it does not print a rate.

What this means for you

When comparing local coding models, note the exact hardware configuration behind each speed figure. If you plan to run Flint locally, expect throughput to depend on memory size and settings, and measure it on your own machine before committing.