October 22, 2026 | AI | By Mahesh

Recursive Superintelligence released Flint on October 21, 2026, its first model, an open-weight coding model with 39.8 billion total parameters and about 2.5 billion active per token under the Apache 2.0 license [1][2][3]. Recursive says it runs at more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 on a 64 GB NVIDIA RTX Spark, both at short context [1]. The company credits its automated research system, the Eureka Machine, with designing much of the stack, and every benchmark and speed figure is self-reported.

What Flint is and what is new

Flint is a text-only coding model designed to run locally, introduced by Recursive chief executive Richard Socher [1][3]. Recursive Superintelligence says its goal is recursively self-improving systems that automate knowledge discovery, and the Eureka Machine is its system for large-scale automated AI research. It proposes ideas, implements them, tests them and uses the results to choose next steps, and Recursive says it contributed to nearly every part of Flint's model-building stack [1]. A small team of people and agents built Flint from scratch in under four months, according to the company [1].

Architecture search started from a baseline of 35 billion total and 3 billion active parameters, and Recursive chose the configuration it projected would balance speed and coding best [1]. It also shipped rec, a lightweight inference engine tuned for Flint on Apple's M4 Pro and M5 Pro and Max chips, which reportedly runs 1.4 to 2.6 times faster than llama.cpp on an M4 Pro [1][2]. The claim that automation built the model has not been independently verified, and RuntimeWire says only future results will show whether the Eureka Machine is more than a compelling account of how Flint was built [3].

Under the hood: experts, attention and quantization

Flint is a decoder-only mixture-of-experts transformer with 49 layers, 2 dense and 47 expert layers. Each expert layer has 256 routed experts, the router picks 6 per token and one shared expert always runs [2]. Thirty-five layers use a 513-token sliding window and 14 use full attention without positional encoding, with 32 query heads and 4 key and value heads. The model card gives a 262,144 token context, while Recursive's blog says up to 512,000 and recommends staying within 256,000 [1][2].

Recursive quantized the output layer to 4 bits, which goes against common practice, and says the NVFP4 build scores within 0.7 points of the 16-bit model. An offline distillation stage with the 16-bit model as teacher nearly halved training steps against a quantization-aware baseline [1]. Checkpoints are 80 GB in BF16, 22.9 GB in NVFP4 GGUF and about 22.6 GB in INT4 for Apple's MLX [2].

Benchmarks: self-reported scores and a mixed comparison

All scores come from Recursive. The BF16 model scores 54.9% on Terminal-Bench 2.1, 74.2% on SWE-bench Verified and 80.3% on LiveCodeBench v6, and the NVFP4 build scores 53.7%, 74.1% and 80.8% on the same tests [2]. Recursive also reports 89.7% on SWE-bench Pro v2, a version it measured itself, so that figure is not comparable with public SWE-Bench Pro results [2]. Terminal-Bench used the Terminus 2 harness, solutions found through web retrieval counted as failures and agentic scores average at least five attempts [2].

The table uses each developer's own model card, and the Terminal-Bench versions differ, so it shows rough standing and not a controlled test [2]. Recursive's own chart puts Flint first among mixture-of-experts models under 50 billion parameters on agentic coding, while dense models such as Qwen 3.8 score higher, at 73.0% on Terminal-Bench 2.1 and 94.9% on SWE-bench Pro v2 [1][2]. Recursive says Flint rivals dense 25 to 40 billion parameter models on coding while running more than five times faster. The speed figures average tokens per second over short contexts, and speed charts use a 128,000 token session [1][2].

Pricing, availability and how to access it

The weights are free under Apache 2.0 on Hugging Face and run with llama.cpp, Ollama, Unsloth, LM Studio, vLLM and SGLang [1]. In Ollama the command is ollama run flint, and the default tag is about 23 GB [4]. MarkTechPost lists a free OpenRouter endpoint with 262,000 tokens of context and tool calling [2].

Hardware needs are modest for a model of this size. Recursive lists Apple's M4 Pro, M5 Pro or M5 Max with 48 GB of memory, or an NVIDIA GPU with at least 48 GB, for the quantized builds, and one B300 or H200 GPU per replica for BF16 [2]. The Ollama page states no memory requirement [4].

Limitations, trade-offs and alternatives

Recursive says Flint is very capable but may need more explicit guidance than larger closed coding models, and it advises splitting complex work into smaller subtasks and staying within 256,000 tokens [1]. Agentic tool use at 2.5 billion active parameters is a trade-off against dense models, and no independent lab has published results [2][3].

For other open-weight coding options, Depth Grid has covered Mistral Large 4, which is far larger and needs server hardware. For a low-cost closed model, see Claude Haiku 5.5 Launch: $0.10 Input, 1M Context, Benchmarks. Teams that care about privacy or offline use should test Flint on their own repositories before relying on the published scores.

Sources

1. Recursive. "Introducing Recursive Flint," October 21, 2026. recursive.com
2. MarkTechPost. "Recursive Releases Flint, an Open-Weight 40B MoE Local Coding Model for Agents with 2.5B Active Params," October 21, 2026. marktechpost.com
3. RuntimeWire. "Recursive ships Flint, a local coding model built with automated AI research." runtimewire.com
4. Ollama. "flint," model library. ollama.com

Read More on Depth Grid

• Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing
• Claude Haiku 5.5 Launch: $0.10 Input, 1M Context, Benchmarks
• Meta Muse Wonder Launch: 1M Context, $4 API Price, Benchmarks
• Nvidia in Talks to Deepen Stake in or Buy Reflection AI

Article by Mahesh | Depth Grid