For months Qwen3.6-35B-A3B has been the reference point for anyone wanting a coding agent to run locally without a GPU farm. Now a new model, Recursive Flint, arrives with numbers that overturn the hierarchy: 40 billion total parameters, 2.5 billion active per token, a 262k token context, and attention split 3:1 between sliding window and full attention. On TerminalBench 2.1 and SWE-bench Pro v2, according to data published in the Hugging Face repository, it clearly beats Qwen3.6-35B-A3B and ties or improves on the other benchmarks shown.

For anyone working with modest hardware, the combination is crucial. Few active parameters lower the computational cost of every inference; the sliding window limits memory growth as the context lengthens, a typical problem for agentic workflows that involve terminals, file systems and multiple rounds of tool calling. The announced support in llama.cpp and the GGUF files prepared by Unsloth signal that the self-hosted community is aiming to run it on consumer GPUs, where every gigabyte of VRAM matters.

The thesis that emerges is not just 'another model'. It is that the next generation of local coding agents will not be driven by larger dense models, but by sparse architectures with hybrid attention and very long contexts, optimised for interactive tasks rather than isolated code completion. The cited benchmarks — TerminalBench and SWE-bench Pro — measure exactly the ability to operate in a real environment, not to answer theoretical questions. If the numbers hold up under independent verification, the gap between cloud assistants and self-hosted solutions will narrow not only on quality but on cost per completed task.

The winners are teams that do not want or cannot send their code to third-party APIs: keeping the agent on internal infrastructure becomes more realistic when the model consumes fewer resources for each reasoning step. The losers, at least in the mid-range, are cloud services selling coding agents based on proprietary LLMs; open and quantizable competition erodes their marginal cost advantage. Qwen3.6-35B-A3B, a venerable model, risks losing its default status for local coding.

For those evaluating on-premise deployment, there are trade-offs between hardware cost, maintenance and data control. AI-RADAR publishes analytical frameworks at /llm-onpremise to navigate these choices without oversimplification.

One note of caution remains: the author of the post is still downloading the model and promises their own benchmarks in the coming days. Until then, the tables in the HF repository are manufacturer claims, not independent truths. But if even part of the results is confirmed, the question would no longer be whether a local coding agent can compete with the cloud, but how far it can go with hardware that sits under a desk and a domestic electricity bill.