Recursive Releases Flint, a 40B Local Coding Model With 2.5B Active Parameters

Flint is available now as an open-weight coding model

Recursive released Flint on October 21, 2026 as a local model for agentic coding. The live RecursiveSI/flint repository exposes the BF16 checkpoint, configuration, tokenizer, chat template and an SGLang serving kit under Apache 2.0. Separate repositories provide quantized artifacts, including a 26 GB Q4_1 GGUF. This is usable artifact availability, not a future open-weight promise.

Flint is designed to investigate repositories, fix bugs, implement features and write tests. Recursive documents local runtimes including rec, llama.cpp, Ollama, LM Studio and Unsloth, GPU serving through SGLang and vLLM, and harness connections for OpenCode, Pi, Cline and OpenHands. Hugging Face currently shows no Inference Provider deployment for the base model, and neither Recursive nor NVIDIA publishes a managed-API price.

The model activates a small fraction of its 39.8B parameters

The model card describes a decoder-only mixture-of-experts Transformer with 39.8 billion total parameters and approximately 2.5 billion active per token, including the output head. Its body alone activates about 2.2 billion parameters. The network has 49 layers: two dense layers followed by 47 mixture-of-experts layers, each with 256 routed experts, six selected experts and one always-active shared expert.

Flint accepts and returns text and supports reasoning and multi-turn tool calls. It does not claim image, audio or video input. The stated knowledge cutoff is June 2026. The card warns that structured tool-call grammar does not make actions safe; applications still need permissions, execution boundaries, code review and tests.

The published context limits do not agree

Recursive’s launch page says Flint supports context windows up to 512,000 tokens and recommends staying within 256,000 for best results. The live BF16 model card instead documents a 262,144-token configuration shared by prompt and generated output. Those statements are close at the recommended operating point but materially different as maximum contracts.

Use 262,144 prompt-plus-output tokens as the conservative current artifact limit unless Recursive publishes a checkpoint-specific configuration and validation for 512,000. Capacity is not uniform accuracy: the model card explicitly says long reasoning can consume the output budget and that the configured window does not guarantee consistent performance across lengths or tasks.

Recursive reports strong coding scores, not independent results

The model card reports 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified. Flint runs used temperature 1.0, top-p 0.95 and top-k 20. Agentic results are average pass@1 over at least five attempts; Terminal-Bench used Terminus 2, and the SWE-bench runs used mini-swe-agent. Recursive says web-retrieved Terminal-Bench solutions were counted as unsuccessful.

Those disclosures improve interpretability but do not create independent validation. Recursive also compares Flint with results reported by other developers whose setups may differ. Treat the numbers as release claims and rerun a versioned repository suite with the same checkpoint, quantization, runtime, harness, context and tool policy before choosing it over another coding model.

Local speed depends on the artifact and runtime

Recursive reports more than 125 output 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. It also says Flint remains fast through a 128,000-token session and that specialized kernels improved RTX Spark generation speed by 55% without changing weights. NVIDIA repeats the 55% claim. AccessAllGPT did not reproduce any of these measurements.

The base BF16 repository is about 80 GB and its included SGLang instructions specify one B300 or H200 GPU per replica. The 26 GB INT4 GGUF lowers the storage and memory threshold for local use, while an NVFP4 variant targets NVIDIA hardware. Precision, runtime, prompt length, KV cache, speculative decoding, thermal limits and harness behavior can all change speed and quality, so one launch number is not a deployment forecast.

Eureka Machine claims need separate evaluation

Recursive says its Eureka Machine helped search architecture choices, quantization recipes, data and training methods, infrastructure and hardware-specific kernels. The company says a small human-and-agent team built Flint from scratch in under four months. It describes an offline distillation stage for the four-bit output layer and says the final NVFP4 model scored within 0.7 points of BF16 across its benchmark set.

That is evidence about Recursive’s development narrative, not proof that automated research made every contribution or that the process generalizes. The public model card gives only a high-level training summary—pretraining, mid-training, supervised fine-tuning and reinforcement learning over text, code, instruction-following and agentic tasks. It does not provide a complete dataset inventory, token count, contamination audit, training compute, evaluation logs or independent reproduction.

What AI coding teams should do next

Trial the exact Flint artifact only if local data handling, Apache-2.0 weights or low active-parameter inference matters. Begin with a repository whose expected behavior is known. Pin the checkpoint revision, runtime, chat template and generation settings; run explanation, focused-test, bug-fix and tool-use tasks; and record compile, test, security, latency, memory and human-review outcomes.

Constrain tool permissions and require review for consequential changes. Wait if a managed endpoint, published service terms, verified 512,000-token operation, independent benchmarks or a complete training-data account is mandatory. Reject the deployment when the 26 GB quantized footprint still exceeds the target hardware, when output quality fails at the required context, or when local operation shifts patching and security work onto a team that cannot own it.

Disclosures

AccessAllGPT received no Recursive, NVIDIA or Hugging Face account, model access, hardware, credits, benchmark logs, private briefing, review, payment or compensation for this article. Recursive, NVIDIA and Hugging Face did not sponsor, review or endorse it. AccessAllGPT did not test, score, rank or deploy Flint. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Recursive, NVIDIA, Hugging Face or any named runtime or harness. Publication-wide relationships are listed on the disclosures page.