Skip to content
Sign in
Technology · Global

Recursive says its AI system built most of coding model Flint

Recursive says its Eureka Machine automated most of the development of Flint, an open-weight coding model. The company says Flint performs strongly on coding benchmarks, while its CEO says AI improved nearly every part of the model pipeline.

AI-assisted synthesis · Based on 5 outlets' reporting · Article text read from 2 of 5 outletsHow it's made

View timeline · 6 entries
5 outletsLatest reporting 3 min read
Coverage: Center OnlyL 0 · C 3 · R 0 · 2 unrated
Updat3 graphic: Recursive
Updat3 graphic

What happened

Key facts

  • Recursive says its Eureka Machine developed most of Flint, an open-weight coding model, largely on its own (per Wired, startupfortune.com).
  • Recursive CEO and cofounder Richard Socher said AI handled and improved nearly every part of Flint’s model pipeline (per Wired, startupfortune.com).
  • Socher said the Eureka Machine developed a new base-model training method, model architectures, and code optimized for specific silicon chips (per Wired, startupfortune.com).
  • Recursive says Flint outperforms or leads models in its size class on coding benchmarks (per Wired, startupfortune.com).
  • Recursive released Flint on October 21 (per startupfortune.com).
  • Flint is an open-weight model designed to run locally (per startupfortune.com, cryptobriefing.com).
  • Flint has a 262,000-token context window and is released under an Apache 2.0 license (per startupfortune.com).
  • Socher said Flint runs at over 125 tokens per second on a MacBook Pro with an M4 Pro chip and more than 150 tokens per second on Nvidia's RTX Spark (per startupfortune.com).

The brief

Recursive says its Eureka Machine carried out most of the development of Flint, an open-weight coding model released on October 21. The company describes Flint as a product of AI-led model research, with CEO and cofounder Richard Socher saying that AI handled and improved nearly every part of the model pipeline.

Socher said the system developed a new base-model training method, model architectures and code tailored to specific silicon chips. He also argued that AI development enables a small team to produce a model that performs strongly in its size class.

WIRED reported his claim that this approach was central to the team’s ability to build Flint; that is Socher’s account of the development process, rather than an independently established breakdown of human and machine contributions. Recursive says Flint outperforms models of similar size and architecture on coding tasks.

Startupfortune.com reports that the company says it leads its size class on SWE-Bench Pro, SWE-Bench Verified and Terminal-Bench, benchmarks intended to assess coding performance. These are company-reported results, and the reporting does not provide a separate independent assessment of the scores. The model is designed to run locally and has a 262,000-token context window.

Recursive says Flint is available under an Apache 2.0 license. Socher said it processes more than 125 tokens per second on a MacBook Pro with an M4 Pro chip and exceeds 150 tokens per second on Nvidia’s RTX Spark. Those speed figures are his claims; the sources do not provide independent testing to verify them.

The coverage also places the release in the context of Recursive’s larger effort to use AI to improve AI systems. WIRED focuses on the company’s claim that its system automated the model pipeline and on concerns about AI improving itself. TechCrunch’s headline likewise describes Flint as largely built by AI, while Crypto Briefing’s headline emphasizes local use and speed.

The latter headline does not supply further performance details beyond characterizing the model as fast. Experts quoted by WIRED warn that systems able to improve themselves could create risks, including potentially escaping human control.

Recursive, by contrast, presents the work as evidence that the steps needed for such a system are in place and argues that AI-led development can help a small team produce capable models.

The central question raised by the release is therefore both technical and governance-related: how much of a model’s development can be automated, and how to assess the risks as that capability advances.

Why it matters

  • Coding-model users are the immediate users of Flint’s benchmark performance and local-running design; Recursive claims it leads its size class on three coding tests.
  • Users running the model on the specified hardware have a concrete performance claim to assess: Socher reported more than 125 tokens per second on a MacBook Pro with an M4 Pro chip and over 150 on Nvidia’s RTX Spark.
  • Recursive’s claimed use of AI across nearly every part of Flint’s pipeline makes the degree of human oversight a practical stake; WIRED quoted experts warning that AI systems improving themselves could potentially escape human control.

What to watch next

  • Whether independent testing supports Recursive’s claims about Flint’s performance on SWE-Bench Pro, SWE-Bench Verified and Terminal-Bench.
  • Whether independent tests reproduce Socher’s reported speeds on the MacBook Pro with an M4 Pro chip and Nvidia’s RTX Spark.
  • How Recursive describes human oversight and safeguards as it develops the Eureka Machine.

How it's being covered

Coverage analysisShow details
5 source articles · 5 unique outlets
Outlet political mix: Center Only· 3 Center · 2 unrated
Where the outlets are based: US/Canada 80% · Global 20%
Outlet political mix
Center Only
Left
Center
Right
0 outlets
3 outlets
0 outlets
The marker shows coverage balance: each unique outlet counts once. 2 unrated outlets are not included.
Geography
Where the outlets covering this story are based.
5 unique outletsDominant coverage: US/Canada
Coverage imbalance
  • Coverage is currently concentrated among Center-rated outlets (3 of 3 rated sources); comparatively few Left or Right-rated outlets have published on the event.
  • Coverage is currently concentrated among outlets based in US/Canada (4 of 5 geographically-identified sources).
How we measure coverage

How outlets' accounts compare

2 framing differences · 2 uncorroborated claims · 3 information gaps
The reports repeat the same account with different emphasis; its central claims are not independently confirmed in the reporting Updat3 read.Compared across 2 reports whose article text Updat3 read; 3 other outlets were checked by headline only.
Framing / emphasis differences
Sources emphasise different angles, actors or consequences of the same reporting.
  • WIRED foregrounds the claim that AI automated most of Flint’s development and includes expert warnings about risks from AI improving itself.
  • Startupfortune.com foregrounds the model’s benchmarks, context window, license and reported hardware speeds; TechCrunch’s headline also describes Flint as largely built by AI, while Crypto Briefing’s headline emphasizes local use and speed.
Updat3: These are differences of emphasis, not disputes about what is reported. They do not show that the reported claims are true.
One-sided or uncorroborated claims
Claims that come from one party or one source family and are not independently corroborated.
  • Richard Socher said AI handled and improved nearly every part of Flint’s model pipeline.
  • Experts quoted by WIRED warned that AI improving itself could potentially escape human control.
Updat3: Not independently corroborated yet — treat as a claim, not an established fact.
Not established by the reporting
Information the available reporting does not contain or confirm. This is a gap, not a disagreement.
  • The sources do not quantify how much of Flint’s work was performed by humans versus the Eureka Machine.
  • The sources do not provide independent benchmark results or independent hardware-speed tests.
  • Recursive’s benchmark leadership and hardware-speed figures are company claims; the sources do not provide independent validation.
Updat3: Updat3 cannot conclude anything about this from the current sources.

Timeline

Developments from this story's sources, dated background we compiled and fact-checked, and our earlier coverage. Each row says which.
  1. Google researchers introduced the transformer architecture in the paper “Attention Is All You Need.”
    Background · compiled and fact-checked by Updat3
  2. Later systems combined transformer architectures with large-scale training on text and source code, then measured coding performance with benchmarks such as HumanEval, published by OpenAI in 2021.
    Background · from our historical context
  3. By 2023, open-model projects such as StarCoder also showed how collaborative datasets and training efforts could produce code-focused models.
    Background · from our historical context
  4. Meta released Code Llama, a family of code-generation models.
    Background · compiled and fact-checked by Updat3
  5. Apple announced 14- and 16-inch MacBook Pro models with M4-series chips.
    Background · compiled and fact-checked by Updat3
  6. Recursive released Flint on October 21.
    Reported in this story's sourcesSource:startupfortune.com

How we got here

OpenAI introduced Codex in 2021, a model trained on code that powered the first version of GitHub Copilot; Meta released Code Llama in 2023, and the BigCode project released StarCoder that year.

The underlying architecture came from the transformer, introduced by Google researchers in the 2017 paper “Attention Is All You Need.” Later systems combined transformer architectures with large-scale training on text and source code, then measured coding performance with benchmarks such as HumanEval, published by OpenAI in 2021.

Original sources

4 of 5 articles · US/Canada · Show all

Every story links to the reporting it was built from. How we choose sources

Spotted a factual error, a broken link or a misleading summary? Tell us.