A wave of model switches is moving through the local LLM practitioner community. On X, developer David Hendrickson (@TeksEdge) posted a direct recommendation at the end of June 2026: "Switch your local model right now", calling out NVIDIA's Nemotron-Cascade-2-30B-A3B as the replacement for Qwen3.5-35B-A3B. The cited reason: "50 tps w/100K context window", roughly 50 tokens per second at a 100,000-token context, faster than most comparable Mixture-of-Experts alternatives at similar memory footprints.
The response from the local AI community has been swift. Nemotron Cascade 2 is increasingly mentioned as the go-to model for local coding agents in June 2026.
What Is Nemotron Cascade 2?
NVIDIA published the Nemotron Cascade 2 research paper in March 2026. The full model name, Nemotron-Cascade-2-30B-A3B, encodes the architecture directly: 30B total parameters, A3B = 3 billion active parameters per forward pass.
That active-parameter ratio is the efficiency story. A conventional 7B dense model uses all 7 billion parameters for every token. Cascade 2's Mixture-of-Experts design activates roughly 10 percent per step, keeping inference fast and memory pressure low while retaining the representational capacity of a 30B model.
The "Cascade" name refers to NVIDIA's post-training alignment methodology: a staged training approach the researchers report produces measurable gains in code reasoning, mathematics, and instruction-following that standard supervised fine-tuning cannot match at this scale.
Not to be confused with Nemotron 3 Ultra (covered here on 27 June): Ultra is the 550B A55B enterprise model designed for server clusters with hundreds of gigabytes of VRAM. Cascade 2 runs on hardware that fits in an office.
Benchmarks: Why Practitioners Are Choosing It
According to community-reported benchmark results, Nemotron-Cascade-2-30B-A3B outperforms both Qwen3.5-35B-A3B and the significantly larger Nemotron-3-Super-120B-A12B across four evaluation categories: mathematics, code reasoning, alignment, and instruction following.
Put plainly: a lighter model with 3B active parameters is beating a model with 4× more total parameters on the tasks that matter most for coding agents. This is what the Cascade training method is designed to achieve.
A side-by-side view of the current local coding agent landscape, based on practitioner measurements:
| Model | Total | Active | ~tok/s | Context |
|---|---|---|---|---|
| Nemotron Cascade 2 | 30B | 3B | ~50 | 100K |
| Qwen3.5-35B-A3B | 35B | 3B | ~35-45 | 32K |
| Qwen3.6 27B MTP | 27B | 27B | ~20-30 | 128K |
| Nemotron 3 Ultra | 550B | 55B | ~8-15 | 1M |
Community-reported figures. Actual throughput varies by hardware, quantisation level, and prompt length.
The 100K context window is worth highlighting separately. Most strong local coding models top out at 32K, 64K tokens, which limits their ability to process complete repositories in a single session. Cascade 2's 100K window covers meaningful real-world codebases without chunking workarounds.
Hardware Requirements
Based on practitioner reports, running Nemotron Cascade 2 at full speed requires:
- Mac Studio M3 Max (96 GB unified) or M3 Ultra (192 GB): The preferred hardware in the community, unified memory eliminates VRAM bottlenecks, sustained throughput at ~50 tok/s with the 4-bit quantised model
- NVIDIA RTX 4090 / RTX 5090 (24 GB VRAM): Q4KM quantised weights fit comfortably; full inference speed at ~50 tok/s reported
- AMD Radeon RX 9070 XT (16 GB VRAM): Q3, Q4 quantisation required; some quality trade-off, but usable for interactive coding assistance
At 4-bit quantisation the model weights occupy approximately 18-22 GB. Practitioners report minimal quality loss compared to the full-precision model on coding tasks, which is the main use case here.
Getting It Running
The recommended path from the community:
- Download GGUF weights: Multiple quantisation options (Q4KM, Q5KM, Q8_0) are available from Hugging Face, official NVIDIA builds and community-converted variants. Match the quantisation to your VRAM budget.
- Start a local inference server: llama.cpp or Ollama both support GGUF natively and expose an OpenAI-compatible API endpoint at
http://localhost:11434/v1(Ollama default) - Connect your IDE plugin: Continue.dev, Cursor, Tabby, and similar tools connect directly to the local endpoint. No token, no account, no cloud billing.
- Set context appropriately: With 100K tokens, you can pass entire directory trees as context. The community recommends structured system prompts that tell the model which files are authoritative to avoid attention dilution.
Zero data leaves your infrastructure. Every prompt, every response, every line of code stays on your hardware.
EU AI Act and GDPR: Local Inference as the Clean Path
Running a coding agent locally on your own server eliminates the entire class of data-transfer compliance risks that come with cloud API integrations:
- No Article 44 GDPR transfer obligations (no data leaving the EU)
- No third-party data processor agreement required for the model itself
- Full auditability, every AI query can be logged and retained internally
- No vendor-imposed data retention or model training on your inputs
Under our reading of the EU AI Act's deployer obligations (Article 26), fully on-premise deployments of general-purpose models significantly reduce the documentation burden compared to cloud API integrations, particularly for sectors handling personal data. This does not constitute legal advice; organisations should verify their specific classification under Annex III.
For European SMBs evaluating local AI strategies, this compliance simplicity is increasingly a primary factor, not a secondary one.
Business Case: Cost Comparison
Cloud-based coding assistants are typically priced at €10-30 per developer per month. For a five-developer team that is €600-1,800 per year, with no control over what data the provider trains on and no guarantee that code stays within EU borders.
A Mac Studio M3 Max (approximately €3,200-4,200) running Nemotron Cascade 2 locally breaks even on pure subscription cost within 24-36 months, by practitioner estimates. The compliance and data control advantages do not appear in that calculation.
EU member states offer various digitisation funding instruments that may cover hardware purchases of this type. Spanish SMBs should check the Kit Digital programme specifically; German businesses can review KfW digitisation loan schemes. Freshlab's pilot project programme covers initial scoping at no cost.
Practical Use Cases
Where Nemotron Cascade 2 delivers the most visible value:
- Whole-repository code review: 100K context means the model sees the actual codebase, not isolated fragments
- Documentation generation: API references, inline comments, and README files written at senior-engineer quality
- Natural-language ERP queries: Local SQL generation without a cloud intermediary (→ /local-ai.html)
- Debugging with full context: Stack traces analysed alongside the relevant source files, no copy-paste into external tools
- Automated test writing: Unit and integration tests generated from existing function signatures, ready to run immediately
Teams already using kAIra Tools can configure Nemotron Cascade 2 as the primary backbone model for the integrated coding assistant.
The Bottom Line
Nemotron-Cascade-2-30B-A3B is the most capable local coding model on consumer hardware as of late June 2026, according to community benchmarks. The combination of ~50 tok/s throughput, 100,000-token context, and MoE efficiency outperforms heavier alternatives on the tasks that matter for real engineering work.
If your team is running Qwen3.5-35B or a comparable model as your local coding agent, the switch is worth making.
Start a pilot project →, Freshlab sets up local AI infrastructure, including coding agent integration, for EU-based SMBs.