What is FrogNano-4B-2609?
FrogNano is a compact repository-level coding agent from Microsoft, released on Hugging Face on 22 September 2026 with a technical report on arXiv. It is Qwen3.5-4B (4.66 billion parameters, a hybrid Gated DeltaNet and gated-attention design) post-trained purely with reinforcement learning on about 1,500 synthetic software-engineering tasks. A pipeline called TaskPilot generated those tasks from real repositories and kept only ones the current model could sometimes but not always solve. No trajectories, reasoning traces or patches from stronger models were used as targets.
It is built to drive Leaf, a five-tool harness (Read, Write, Edit, Glob and Bash) that runs in Kubernetes sandboxes and is published on GitHub. It reads and writes text only; the image and video parts inherited from Qwen3.5-4B were not trained or evaluated. The evaluated setup uses about 131K tokens of context, 150 agent steps and up to 8,192 generated tokens per turn.
Microsoft describes it as a research artefact, not a production model. The Hugging Face repository is tagged MIT while the model card's summary table says Apache License 2.0; the Leaf harness on GitHub is MIT.
Is FrogNano-4B-2609 worth it?
FrogNano's numbers are striking for its size. Inside the Leaf harness, averaged over three seeds, Microsoft reports 61.5% on SWE-bench Verified, up from 39.4% for the untouched Qwen3.5-4B, plus 37.6% on SWE-bench Pro, 31.1% on Terminal-Bench 2.0 and 47.3% on PatchEval-Verified, a set of 230 real security-fix cases. With three attempts per task, at least one succeeds 71.0% of the time on Verified. The paper says this is competitive with models six to eight times its size on SWE-bench Pro at a quarter of the cost. No independent leaderboard has run it yet.
Read those numbers with the method in mind. SWE-bench Verified was the validation set used to choose checkpoints across five training rounds, so the held-out figures (Pro, Terminal-Bench, PatchEval) are the fairer guide. Training data was Python-heavy and English, and the card says performance is sensitive to the harness: outside Leaf, expect less. The paper reports no successful reward hacks, while noting that some attempts to weaken tests were blocked only by the sandbox.
The interesting part is the recipe, not the checkpoint. Microsoft shows that about 300 synthetic tasks per round, generated against the model's own failures, matched a baseline of real tasks that had been filtered through costly rollouts. For teams training their own small agents, the report and harness are more useful than the weights.
As a tool to run, it is demanding. The evaluation harness wants Python 3.12, a Kubernetes cluster with pod and network-policy permissions, and an OpenAI-compatible endpoint with the right tool-call parser. The weights need about 9.3 GB in BF16 before the KV cache, and no quantised version has been validated. Microsoft did no safety-specific training and says the model should not be treated as aligned for unattended use.
Where it earns its place
A research engineer at a university lab with one 8-GPU node
Run coding-agent RL experiments without a frontier-model budget
The 4B model, Leaf harness and benchmark configs reproduce Microsoft's setup; the paper's training ran on a single node of eight B200 GPUs. Starting from FrogNano rather than raw Qwen3.5-4B gives a 22-point head start on SWE-bench Verified.
A platform engineer at a 200-developer Python shop with strict code-egress rules
Draft fixes for small, well-tested bugs without sending code to an API
Serve the weights with SGLang inside the network and point Leaf at an isolated repository copy with network access off; it proposes a patch and runs the tests. Treat every patch as a draft: the card warns patches can pass tests yet be wrong or insecure.
Before you commit
- Setup
- Download the weights and serve them on a runtime that supports Qwen3.5 (SGLang is the documented one, with --reasoning-parser qwen3 and --tool-call-parser qwen3_coder). The Leaf harness installs with uv or pip on Python 3.12 and needs a Kubernetes namespace for its sandboxes.
- Integrations
- Leaf talks to any OpenAI-compatible endpoint and exposes five tools: Read, Write, Edit, Glob and Bash. Presets cover SWE-bench Verified, SWE-bench Pro, Terminal-Bench 2.0 and PatchEval, with optional Weights & Biases tracking.
- Data and privacy
- Self-hosted, so code stays on your infrastructure. The model card asks for a least-privilege sandbox with network access off by default; FrogNano received no safety, refusal or adversarial training of its own.
- Cost at scale
- No API price. BF16 weights take about 9.3 GB plus KV cache, which grows substantially near the 131K-token limit; CPU, desktop, mobile and quantised use have not been validated. Microsoft's benchmark presets give each task up to 10,800 seconds and 150 steps.
- Limits
- About 131K tokens of context and 8,192 generated tokens per turn as evaluated; English and Python are the only validated languages. Image and video inputs are unsupported despite inherited components.
What’s good
- 61.5% SWE-bench Verified, 37.6% SWE-bench Pro at 4B (self-reported)
- No distillation from larger models; recipe fully described
- About 9.3 GB of BF16 weights before KV cache
- Leaf harness and evaluation configs open on GitHub
What’s not
- No independent benchmark yet; Verified was the validation set
- Python-heavy, English-only training; harness-sensitive
- Evaluation harness needs a Kubernetes cluster
- Licence stated as both MIT and Apache 2.0
Best for
Researchers and platform teams studying small coding agents, or running sandboxed bug-fix agents on their own GPUs where every patch is reviewed by a person.
FrogNano-4B-2609 alternatives
FrogNano-4B-2609: common questions
- Is FrogNano-4B-2609 worth it?
- A 4B model that resolves six in ten SWE-bench Verified issues is a real research result, but every number is Microsoft's own and the harness is built for a lab, not a laptop. We rate it 3.7 out of 5.
- How much does FrogNano-4B-2609 cost?
- Free weights (tagged MIT on Hugging Face; the card's table says Apache 2.0). No hosted API; you serve it yourself, for example with SGLang.
- What are the downsides of FrogNano-4B-2609?
- No independent benchmark yet; Verified was the validation set; python-heavy, English-only training; harness-sensitive; evaluation harness needs a Kubernetes cluster; licence stated as both MIT and Apache 2.0.
- Who is FrogNano-4B-2609 best for?
- Researchers and platform teams studying small coding agents, or running sandboxed bug-fix agents on their own GPUs where every patch is reviewed by a person.
- How much does FrogNano-4B-2609 cost for a team or in production?
- No API price. BF16 weights take about 9.3 GB plus KV cache, which grows substantially near the 131K-token limit; CPU, desktop, mobile and quantised use have not been validated. Microsoft's benchmark presets give each task up to 10,800 seconds and 150 steps.
- Is FrogNano-4B-2609 safe for company data?
- Self-hosted, so code stays on your infrastructure. The model card asks for a least-privilege sandbox with network access off by default; FrogNano received no safety, refusal or adversarial training of its own.
- What are the best FrogNano-4B-2609 alternatives?
- The closest alternatives we have reviewed are Devstral Small 2, gpt-oss-20b, LFM3 and Devstral 2.
Answers checked Oct 31, 2026.