FrogNano gets a 3.48 GB community GGUF build for local coding
A new mixed-precision conversion brings Microsoft's compact coding model to Ollama and llama.cpp, with an author-reported perplexity comparison and clear limits on what it proves.
A smaller route into Microsoft’s coding model
Community maintainer roman220220 has uploaded a 3.48 GB GGUF conversion of FrogNano for Ollama and llama.cpp. The October 17 build uses an importance matrix and different precisions for different parts of the model, giving local users a concrete download and documented launch commands.
This is a new community conversion, not a new Microsoft model launch. Microsoft’s English model card dates FrogNano to September 22. It describes a compact repository-level coding model derived from Qwen3.5-4B and post-trained with reinforcement learning on approximately 1,500 synthetic software-engineering tasks.
What the mixed-precision recipe changes
The author keeps full-attention tensors and token embeddings at Q8_0, puts the gated-DeltaNet linear-attention tensors at Q6_K, and uses the Q4_K_M base preset for feed-forward tensors. The card says its importance matrix used 200 chunks of bartowski’s calibration corpus on an RTX 4090.
The published file listing contains a 3,477,677,760-byte GGUF. The author compares it with an 8.67 GB F16 conversion and reports WikiText-2 perplexity of 8.703 versus 8.623, a 0.93% increase, measured over 40 chunks of 512 tokens. That is the author’s language-modelling comparison, not a coding-agent benchmark or an AITrending reproduction.
Microsoft reports 61.5% Avg@3 on SWE-bench Verified in its Leaf harness for the original model. That result must not be transferred to this quantization: the community author explicitly says the coding benchmarks were not rerun on the GGUF build. File size also excludes runtime state and context memory, so 3.48 GB is not a verified total RAM or VRAM requirement.
How local users can try it
The card documents an Ollama command:
ollama run hf.co/roman220220/FrogNano-4B-2609-GGUF-jang-imatrixFor llama.cpp, it gives llama-server -m frognano-4b-jang.gguf -ngl 99 --jinja --port 8080 after downloading the file. The author recommends applying the model’s own chat template and using temperature 0.6. No paid inference API credential is specified for these local commands; users supply compatible hardware, storage and the inference runtime. The card does not pin a minimum runtime version.
Serving the weights alone does not install a complete repository agent. Microsoft’s evaluated workflow uses the separate Leaf harness for file inspection, editing, search and shell execution in an isolated environment. Its outputs are candidate patches for human review and testing, rather than automatically approved deployments.
This GGUF is text-only. The card says it omits the vision tower and MTP speculative-decoding head; Microsoft’s own card does not claim evaluated image or video support for FrogNano. The community card declares Apache-2.0, matching the base card’s license section, although Microsoft’s YAML metadata says MIT. That inconsistency remains visible rather than being treated as a newly clarified license.
