---

license: mit

base_model: microsoft/FrogNano-4B-2609

tags:

• gguf
• llama.cpp
• qwen3_5
• code
• coding-agent

---

FrogNano-4B-2609 — GGUF quants for llama.cpp

This repo contains llama.cpp GGUF conversions of microsoft/FrogNano-4B-2609 — Microsoft Corporation's compact repository-level coding agent (released 22-SEP-2026). All 29 standard quant types are here, each cut from the BF16 checkpoint with an importance matrix computed on those same BF16 weights, plus the BF16 file itself and the vision-tower export, quantized 2026-10-04.

These files are llama.cpp GGUF. They are not loaded by vLLM, SGLang or Transformers the way the source repo's safetensors are — use llama-cli, llama-server, or another GGUF runtime.

Quick picks

• Most people: Q4_K_M (2.59 GB) — the usual default, best quality-per-byte.
• Near-lossless: Q6_K (3.32 GB), or Q8_0 (4.29 GB) if RAM is not a constraint.
• Tight on memory: IQ4_XS (2.40 GB) is usually the best sub-4.5 bpw option; IQ2_M (1.69 GB) is the smallest size most people find usable.
• Curious about the floor: IQ1_S (1.33 GB) / Q1_0 (1.03 GB) — expect substantial degradation; included for completeness.

What is FrogNano?

From Microsoft's model card: FrogNano is a repository-level coding agent derived from Qwen/Qwen3.5-4B, post-trained exclusively with reinforcement learning on ~1,500 synthetic software-engineering tasks that were generated, validated, and calibrated against the evolving policy using TaskPilot. It operates through the lightweight five-tool Leaf harness (file read/write/edit, glob-based search, shell execution) inside an isolated repository sandbox. A distinctive choice: the agent post-training used no stronger-model solution trajectories, actions, reasoning traces, or patch targets — the model learned from its own policy's trajectories, rewarded by executable test results.

Architecture (from the source card and config): a dense 32-layer hybrid Gated DeltaNet / gated-attention causal LM (Qwen3_5 family), ~4.66B parameters, hidden size 2,560, vocabulary 248,320.

Its evaluated configuration is an agent loop: ~131K tokens of combined context with a budget of 150 interaction steps in the Leaf harness, and up to 8,192 generated tokens per assistant turn. Supported interaction is English natural language over Python-heavy code; a response containing tool calls continues the interaction, one without them ends it.

Files

| File | Size | Notes |

|---|---|---|

| FrogNano-4B-2609-BF16.gguf | 8.07 GB | Full-precision source. Every quant below is cut from this file. |

| mmproj-FrogNano-4B-2609-F16.gguf | 0.63 GB | Vision-tower export — see the text-only note below. |

| FrogNano-4B-2609-Q8_0.gguf | 4.29 GB | Effectively lossless. Use when disk and RAM are not the constraint. |

| FrogNano-4B-2609-Q6_K.gguf | 3.32 GB | Near-lossless; the last stop before quality becomes measurable. |

| FrogNano-4B-2609-Q5_K_M.gguf | 2.94 GB | Very good quality, noticeably smaller than Q6_K. |

| FrogNano-4B-2609-Q5_1.gguf | 3.08 GB | Legacy. Prefer Q5_K_M. |

| FrogNano-4B-2609-Q5_K_S.gguf | 2.86 GB | Slightly smaller than Q5_K_M for a slight quality cost. |

| FrogNano-4B-2609-Q5_0.gguf | 2.86 GB | Legacy. Prefer Q5_K_M. |

| FrogNano-4B-2609-Q4_K_M.gguf | 2.59 GB | The usual default. Best quality-per-byte for most people. |

| FrogNano-4B-2609-Q4_K_S.gguf | 2.45 GB | A little smaller than Q4_K_M, a little worse. |

| FrogNano-4B-2609-IQ4_NL.gguf | 2.49 GB | Non-linear 4-bit; good on hardware without fast K-quant kernels. |

| FrogNano-4B-2609-IQ4_XS.gguf | 2.40 GB | Best sub-4.5bpw option; usually beats Q4_K_S at a smaller size. |

| FrogNano-4B-2609-Q4_1.gguf | 2.65 GB | Legacy. Prefer Q4_K_M. |

| FrogNano-4B-2609-Q4_0.gguf | 2.43 GB | Legacy round-to-nearest. Prefer Q4_K_M unless a runtime needs this. |

| FrogNano-4B-2609-Q3_K_L.gguf | 2.32 GB | Small, with real quality loss. Usable when RAM is tight. |

| FrogNano-4B-2609-Q3_K_M.gguf | 2.17 GB | Smaller again; noticeable degradation. |

| FrogNano-4B-2609-IQ3_M.gguf | 2.08 GB | Strong at ~3.7bpw, clearly better than Q3_K_M. |

| FrogNano-4B-2609-IQ3_S.gguf | 2.06 GB | Slightly smaller than IQ3_M. |

| FrogNano-4B-2609-IQ3_XS.gguf | 2.00 GB | Aggressive but coherent. |

| FrogNano-4B-2609-Q3_K_S.gguf | 1.99 GB | Aggressive. Prefer IQ3_M at a similar size. |

| FrogNano-4B-2609-IQ3_XXS.gguf | 1.84 GB | Very aggressive; the last coherent step down. |

| FrogNano-4B-2609-Q2_K.gguf | 1.85 GB | Very small, heavily degraded. For experimentation. |

| FrogNano-4B-2609-Q2_K_S.gguf | 1.79 GB | Smaller than Q2_K, at a further quality cost. |

| FrogNano-4B-2609-IQ2_M.gguf | 1.69 GB | The smallest size most people find usable. |

| FrogNano-4B-2609-IQ2_S.gguf | 1.60 GB | Below the usual usability line. |

| FrogNano-4B-2609-IQ2_XS.gguf | 1.58 GB | Experimental. |

| FrogNano-4B-2609-IQ2_XXS.gguf | 1.50 GB | Experimental. |

| FrogNano-4B-2609-Q2_0.gguf | 1.50 GB | Extreme, group-64. Included for completeness. |

| FrogNano-4B-2609-IQ1_M.gguf | 1.39 GB | Extreme. Expect substantial degradation. |

| FrogNano-4B-2609-IQ1_S.gguf | 1.33 GB | Extreme. Expect substantial degradation. |

| FrogNano-4B-2609-Q1_0.gguf | 1.03 GB | Extreme. Included for completeness. |

Text-only — including the vision file

The upstream Qwen3.5-4B checkpoint carries image and video input components, which is why the converted vision tower is included here (mmproj-FrogNano-4B-2609-F16.gguf, 0.63 GB). Microsoft's card is explicit, though: those modalities were not post-trained or evaluated for FrogNano and are unsupported. FrogNano's intended and evaluated inputs are text only — natural-language instructions, source code, and textual tool outputs. Don't pick this model for multimodal work.

Benchmarks (measured by Microsoft, not by this conversion)

From the source model card, evaluated with the Leaf harness at Avg@3 single-attempt success (temperature 0.6, repetition penalty 1.0, 150 steps, ~131K context):

| Benchmark | FrogNano | |

|---|---|---|

| SWE-bench Verified | 61.5% | Qwen3.5-4B base under the same harness: 39.4%; across five RL iterations: 49.1 → 53.1 → 56.7 → 59.1 → 61.5 |

| SWE-bench Pro | 37.6% | 731 longer-horizon tasks across Python, JavaScript, TypeScript, Go |

| Terminal-Bench 2.0 | 31.1% | 89 tasks, graded on final container state |

| PatchEval-Verified | 47.3% | 230 real CVE repair cases (2015–2025) |

Pass@3 (at least one of three trajectories succeeds): 71.0% on SWE-bench Verified, 47.6% on SWE-bench Pro. These are full-precision source numbers — the quants here were not re-evaluated. FrogNano has not been separately evaluated for general language understanding, mathematics, or multilingual performance, and upstream Qwen3.5-4B scores in those areas should not be attributed to it.

Limitations worth reading before you download

Condensed from the source card:

• Specialized for English, Python-heavy repository tasks; sensitive to the agent harness, prompt format, available tools, and test quality. May underperform on non-English instructions, other languages/frameworks, unfamiliar repo structures, or projects without reliable tests.
• Patches may be incorrect or insecure even when passing available tests. The model proposes changes; a qualified human must review them, plus regression and security testing, before use.
• Not independently safety-aligned: post-training was coding-focused (executable rewards) with no dedicated safety-preference, refusal, or adversarial datasets. Not validated for unrestricted autonomous deployment, safety-critical use, or acting on production systems.
• Inherits potential biases and representation gaps from Qwen3.5-4B; these were not comprehensively reevaluated after the coding-focused RL.

Usage

# set the repo once, then use it everywhere below
repo=NANI-Nithin/FrogNano-4B-2609-GGUF

# pull and run a quant straight from the Hub
llama-cli -hf "$repo":Q4_K_M -p "Hello"

# serve it
llama-server -hf "$repo":Q4_K_M

# or download one quant without running it
huggingface-cli download "$repo" --include "*Q4_K_M*" --local-dir .

After a manual download, point llama-cli -m at the downloaded file. Since FrogNano is an agent model, the interesting use is a tool loop: per the source card, integrating it as an agent requires the tokenizer and chat template, JSON-schema tool-call parsing, and the interaction loop (Microsoft's is Leaf), with tool calls executed in an isolated, least-privilege sandbox.

Licence

The source repository's Hub metadata declares MIT; its model card's summary table states Apache License 2.0, noting FrogNano is derived from Qwen/Qwen3.5-4B with applicable upstream copyright and attribution notices retained. The source card is the authoritative statement — read both before relying on either.

Links

• Source model: microsoft/FrogNano-4B-2609 (Microsoft Corporation)
• Upstream dependency: Qwen/Qwen3.5-4B
• Technical report: aka.ms/frognano-tech-report
• Leaf harness: github.com/microsoft/FrogNano
• Contact (per source card): MSFTAIActRequest@microsoft.com