AI NewsOct 22October 22, 20267 min read

Flint LLM Brings 40B Coding to Local Hardware

Recursive’s open-weight coding model targets offline development, but its context limits, memory requirements, and speed claims need hardware-specific reading.

Listen

AI narration

11:12

AI SummaryGenerated from this article

Recursive released Flint, a 40-billion-parameter open-weight coding model designed to run locally on Apple Silicon and NVIDIA hardware without sending code to hosted services. The mixture-of-experts model activates approximately 2.5 billion parameters per token and supports up to 512K tokens of context, though Recursive recommends staying within 256K for best results. Flint requires at least 48GB of unified memory on supported Apple chips and the BF16 checkpoint is roughly an 80GB download. Recursive reports generation speeds above 125 tokens per second on M4 Pro and above 150 on NVIDIA RTX, and claims leadership on agentic coding benchmarks among models below 50 billion parameters. The company says its Eureka Machine automated research system contributed throughout development, but all performance claims remain vendor-reported and require hardware-specific testing.

Recursive’s optimized Mac setup for Flint requires a supported Apple chip and at least 48GB of unified memory. For developers with that hardware, the 40-billion-parameter coding LLM offers a way to work without sending code to a hosted model.

Recursive’s announcement describes a mixture-of-experts model that activates approximately 2.5 billion parameters per token. The company says it supports up to 512K tokens of context, although it recommends staying within 256K for best results.

The October 21 release carries another claim: Recursive says its Eureka Machine automated research system contributed throughout development, helping a small team build Flint from scratch in under four months.

For developers, the immediate question is whether Flint delivers useful coding assistance on hardware they own. Its benchmark leadership, inference speeds, and automated-development results are vendor-reported. The downloadable weights make those claims testable; they are not independently established results.

Local Coding Requires More Than a Small Active Count

Flint is designed to investigate repositories, fix bugs, implement features, and write tests. Its weights are available under Apache 2.0, according to both the announcement and the Hugging Face model card.

Developers can use it as an alternative to paying a hosted service for model inference, keeping code and prompts on their own machine in a fully local workflow. Hardware, electricity, and maintenance still cost money, but local operation removes the hosted model’s per-token API bill.

Flint has approximately 39.8 billion parameters in total, rounded to 40B in its release branding. Only about 2.5 billion participate in generating each token, reducing the active computation compared with using the entire model every time.

For memory planning, though, Flint is not equivalent to a 2.5B model. The other experts still belong to the model and need storage and a suitable memory-management strategy. The BF16 checkpoint is roughly an 80GB download; the model card distinguishes it from separately prepared quantized variants for NVIDIA and Apple Silicon runtimes.

Recursive’s local setup documentation specifies an M4 Pro, M5 Pro, or M5 Max with at least 48GB of unified memory for its optimized rec Mac engine. That is a much more useful purchasing constraint than the phrase “runs on a laptop.”

Let’s work together

We’re open to thoughtful collaborations with teams building in AI. Explore the ways we can work together.

See collaboration details

Anthropic’s Amodei Defends ‘Race to the Top’ Strategy

Oct 22 • 7 min

Walmart Sets AI Shopping Rules Ahead of Profile Launch

Oct 22 • 6 min

Granite Speech 5.0 Adds Languages and Faster Transcription

Oct 22 • 6 min

Meta Says Muse Tops 10 Million Users as Shopping Stands Out

The Coding Scores Need Their Harnesses Attached

Recursive’s model card reports the following release results:

Terminal-Bench 2.1: 54.9%.

SWE-bench Pro v2: 89.7%.

SWE-bench Verified: 74.2%.

The company says Flint leads mixture-of-experts models below 50 billion total parameters on its highlighted agentic coding benchmarks. This is narrower than claiming leadership across all coding AI models, and it depends on the evaluation conditions.

Recursive used Terminus 2 for Terminal-Bench and mini-swe-agent for the SWE-bench evaluations. Agentic results are reported as average pass@1 over at least five attempts, with agent timeouts of up to five hours. Averaging repeated single-attempt outcomes is not the same as giving an agent five chances and counting any eventual success.

Network conditions differed, too. Terminal-Bench allowed internet access, but Recursive counted solutions obtained through retrieval as unsuccessful. The SWE-bench evaluations ran without internet access.

The model card explicitly warns that comparison setups vary. Recursive measured some comparison scores; others came from the respective model developers, who may not apply the same retrieval adjustment.

The results support taking Flint seriously as a local coding candidate, without establishing a universal ranking across hardware, harnesses, and real repositories. A developer’s own task completion rate, intervention burden, and elapsed time will be more useful than an aggregate leaderboard position.

What Recursive Says Its Research Agents Built

Recursive describes Eureka Machine as a system that proposes ideas, implements them, runs experiments, and uses the results to choose subsequent work. The company says it contributed across Flint’s architecture, data, training recipes, quantization, infrastructure, and inference optimization.

Several reported contributions bear directly on local computing. Recursive says Eureka explored expert configurations starting from a 35B-total, 3B-active baseline. The resulting 40B-total, 2.5B-active design was projected to offer the best speed-and-coding tradeoff among the configurations tested.

The company also says the system selected a four-bit output layer and developed a distillation stage to recover accuracy. Recursive reports that the final NVFP4 model scores within 0.7 points of the 16-bit model across its benchmarks.

On the inference side, Eureka reportedly produced hardware-specific kernels, including a Metal megakernel that fuses the decode forward pass on M4 and M5 chips.

These claims describe proposed changes across several parts of model building, going beyond a general assertion that AI helped write training code. The launch material does not, however, establish how much development time automation saved against a comparable human-led effort.

One AI-assisted release also cannot demonstrate a sustained cycle in which each model reliably improves the research system that builds its successor. Flint tests that thesis without proving that Recursive’s broader self-improvement goal has been achieved.

The nearer-term test is whether developers can reproduce Flint’s performance on the specified local hardware and finish real coding tasks with an acceptable amount of supervision. If so, the release offers value even without validating the larger research narrative. Its strongest proposition is control over where inference happens, paired with enough coding capability to make that control practical.

Flint LLM Brings 40B Coding to Local Hardware

Local Coding Requires More Than a Small Active Count

Tags

Read next

Let’s work together

Related articles

Anthropic’s Amodei Defends ‘Race to the Top’ Strategy

Walmart Sets AI Shopping Rules Ahead of Profile Launch

Granite Speech 5.0 Adds Languages and Faster Transcription

The 512K Context Claim Comes With a 256K Recommendation

Fast Generation Is Not the Same as Fast Task Completion

The Coding Scores Need Their Harnesses Attached

What Recursive Says Its Research Agents Built

Sources

Read next

Tags

Related articles

Anthropic’s Amodei Defends ‘Race to the Top’ Strategy

Walmart Sets AI Shopping Rules Ahead of Profile Launch

Granite Speech 5.0 Adds Languages and Faster Transcription

Meta Says Muse Tops 10 Million Users as Shopping Stands Out

OpenAI Says GPT-5.6 Sol Helped Find Critical PHP Flaw