02BFastest
Nimbus-2B
A compact coding partner for quick edits, local automation, and lightweight agent loops.
- Device target
- 8 GB class
- Release candidate
- Q5_K_M
- Control
- BF16
Nimbus Models · Coming soon
Small models.
Serious work.
A new family of local coding and reasoning models, shaped for the memory people actually have—not a datacenter they rent by the token.
No score ships before the underlying run, protocol, and checksum do.
Release preview
Direct mode · pass@1ModelHumanEval+MBPP+
Nimbus-2BFast · Q5_K_M
46.3%44.7%Nimbus-4BBalanced · Q5_K_M
68.3%61.6%Nimbus-9BDeep · Q5_K_M
81.1%70.9%1,626 scored generationsComplete · checksummed
One family · Four operating points
Choose the model your machine can keep close.
The first three Nimbus models are in release preparation now. Q5_K_M is the consumer release candidate; Nimbus-12B remains in the lab and will follow later.
04BBalanced
Nimbus-4B
The middle path: more planning room and code context while remaining practical on everyday hardware.
- Device target
- 8–16 GB
- Release candidate
- Q5_K_M
- Control
- BF16
09BDeepest
Nimbus-9B
The most capable first-wave Nimbus model, aimed at deeper reasoning and longer multi-step coding work.
- Device target
- 16 GB class
- Release candidate
- Q5_K_M
- Alternate quant
- Q4_K_M
12BIn the lab
Nimbus-12B
The next weather-named Nimbus model, planned for deeper agentic work on high-memory consumer machines.
- Status
- Coming later
- Device target
- 16–32 GB
- Scores
- Not evaluated
Device classes are release targets, not completed compatibility claims. Final recommendations will use measured complete working-set memory—not model file size alone.
Verified artifacts · July 2026
Built for memory you can own.
These are measured GGUF file sizes, not estimated bits per weight. Runtime memory also includes the context cache, compute buffers, and application overhead.
Nimbus-2B8 GB target
Nimbus-4B8–16 GB target
Nimbus-9B16 GB Q5 target
Positioning: Nimbus-9B Q5_K_M targets 16 GB devices. The BF16 reference is best treated as a 24 GB minimum / 32 GB recommended desktop or high-memory laptop artifact.
Complete direct-mode evaluation
Four code scores. Every model. Every task.
All three Q5_K_M release candidates completed the full EvalPlus suite in direct mode: 164 HumanEval tasks and 378 MBPP tasks per model. Base and Plus scores are pass@1.
HumanEval
Canonical Python function-completion correctness.
9B Q5 · 82.3%HumanEval+
Expanded tests designed to catch fragile solutions.
9B Q5 · 81.1%MBPP
Hundreds of short programming tasks from natural-language descriptions.
9B Q5 · 83.3%MBPP+
A stricter expanded test suite for the same task family.
9B Q5 · 70.9%Q5 direct mode · Verified
Capability that scales with the footprint.
Expanded-test pass@1 rises sharply across the lineup. The 9B release candidate reaches 81.1% on HumanEval+ while remaining a 6.02 GiB model file.
| Artifact | Role | HumanEval | HumanEval+ | MBPP | MBPP+ |
|---|---|---|---|---|---|
| Nimbus-2B Q5_K_M | Coming soon | 50.0% | 46.3% | 53.4% | 44.7% |
| Nimbus-4B Q5_K_M | Coming soon | 73.8% | 68.3% | 75.4% | 61.6% |
| Nimbus-9B Q5_K_M | Coming soon | 82.3% | 81.1% | 83.3% | 70.9% |
Recent compact-model context
A strong first showing.
Officially reported results from recent compact releases provide useful context, but are directional—not a shared leaderboard. Prompt formats, shot counts, runtimes, and model precision differ.
Nimbus-4B 73.8%
Nimbus-4B 68.3%
Nimbus-9B 82.3%
Nimbus-9B 81.1%
Sources: IBM Granite 4.1 official model card and Google Gemma 3 official model card. Nimbus scores use our complete direct-mode EvalPlus Q5_K_M runs; third-party scores use each publisher's stated protocol.
Current public protocol
Direct answers. Complete coverage.
The first published scores use deterministic direct mode. Separate native-thinking experiments remain research results and are not mixed into this chart.
The model returns implementation code without a hidden reasoning pass.
The same consumer release-candidate format across all three rows.
Repeatable generations without temperature variance.
Pinned generation and EvalPlus revisions with sample checksums.
Exact task count required before scoring.
Exact task count required before scoring.
Nimbus Labs · First model family
Coming soon.
Nimbus-2B, Nimbus-4B, and Nimbus-9B are being packaged now. Model files, detailed cards, manifests, and checksums will arrive together; Nimbus-12B follows later.