A deeper local reasoning model for difficult coding, debugging, and tool-aware problem solving.
Nimbus is a family of local coding models: 2B for speed, 4B for balance, and 9B v2.1 for deeper reasoning. This Transformers repository contains the merged BF16 checkpoint and full multimodal components. The corresponding GGUF repository is text-only.
The adjacent assets/nimbus-family-footprint.json contains the plotted values. Download size is not runtime memory: context cache and runtime buffers require additional capacity.
The released Q5_K_M artifact was evaluated with llama.cpp b10007 on Vulkan, native thinking, one answer per task, temperature 0.6, top-p 0.95, top-k 20, seed 42, and the official network-disabled EvalPlus scorer. The base run allowed 32,768 completion tokens. Each original length-plus-empty nonresponse received exactly one preserved 60,000-token recovery attempt; there was no repeated resampling.
Benchmark
Passed
Total
pass@1
HumanEval
146
164
89.0%
HumanEval+
135
164
82.3%
MBPP
330
378
87.3%
MBPP+
277
378
73.3%
Three HumanEval and seven MBPP truncations recovered at the larger limit. Two HumanEval and three MBPP nonresponses remained zero-valued. “Plus” uses EvalPlus' official base-and-plus pass condition.
Nimbus-9B v2.1 EvalPlus results
The adjacent assets/nimbus-9b-v2.1-evalplus.json is the machine-readable source for this chart.
Release landscape
This is a capability and packaging comparison, not a cross-model score claim. Other publishers' benchmarks use different harnesses, prompts, inference budgets, and hardware, so they are not mixed into the Nimbus EvalPlus chart.
Release
Parameters
Modality
Published context
Positioning
Nimbus-9B v2.1
9B
Multimodal Transformers; text-only GGUF
262,144 validated for this release
Local coding/reasoning; BF16, Q4_K_M, Q5_K_M with checksums
General instruction, code, function calling, and RAG
Nimbus-9B v2.1 inherits its architecture and foundation capabilities from Qwen3.5 through Ornith. Nimbus' contribution is the post-training, release packaging, quantization, and exact evaluation reported above—not authorship of the foundation model.
Quickstart
Use a recent Transformers release compatible with Qwen3.5. Load the repository with trust_remote_code=False, preserve the supplied processor/tokenizer files, and enable thinking through the supplied chat template where supported. Validate generated code before execution.
Relationship: Nimbus post-training / fine-tune and verified merge
Exact uploaded file hashes: SHA256SUMS
Release metadata: release-manifest.json
Nimbus prominently credits Qwen3.5 as the foundation for this model family. Nimbus-9B-v2.1 additionally credits DeepReinforce's Ornith-1.0-9B as its immediate upstream.
Intended use
Local coding assistance, code explanation, debugging, test generation, and bounded tool-aware workflows. Host applications must enforce permissions, sandboxing, timeouts, and verification.
Limitations
The model can produce incorrect, insecure, incomplete, or non-compiling code. Benchmark performance does not guarantee project-level correctness. GGUF artifacts are text-only even though this Transformers checkpoint includes multimodal components.
Licenses and notices
See LICENSES.md, THIRD_PARTY_NOTICES.md, and the bundled Apache-2.0 text. The model is a derivative distribution; Nimbus attribution does not replace upstream attribution.
