The latest local AI models in 2026: Kimi K3, Inkling, Laguna, DeepSeek, Qwen, Muse and Nemotron

The short answer

There is no single local winner. Kimi K3 and Qwen3.8-Max target frontier capability but need server clusters. Inkling and Inkling-Small prioritize multimodality and customization. Laguna S 2.1 and DeepSeek V4 Flash focus on agentic coding. Muse Glimmer targets powerful consumer GPUs, while Nemotron 3.5 Lightning prioritizes speed and throughput.

• Highest capability: Kimi K3 and Qwen3.8-Max.
• Agentic coding balance: Laguna S 2.1 and DeepSeek V4 Flash 0731.
• Multimodal customization: Inkling-Small.
• Consumer hardware: Muse Glimmer when its quantization fits your VRAM.
• High throughput: Nemotron 3.5 Lightning.

Kimi K3: open frontier, not a laptop model

Moonshot released Kimi K3 with 2.8T total parameters, about 104B active parameters, a one-million-token context and native multimodality. Published scores such as 93.5 on GPQA Diamond and 88.3 on Terminal-Bench 2.1 put it close to closed frontier systems. Its practical limit is infrastructure: even MXFP4 deployment needs distributed hardware and roughly 1.5 TB for weights plus operational overhead.

Inkling and Inkling-Small: open multimodality

Thinking Machines Lab released Inkling, a 975B-total and 41B-active MoE, followed by Inkling-Small at 276B total and 12B active. Both accept text, image and audio, support up to one million tokens and use Apache 2.0. Small matching its larger sibling on some tasks is a useful warning against selecting models by parameter count alone.

Laguna S 2.1 and DeepSeek V4 Flash for coding agents

Laguna S 2.1 combines 117.6B total parameters with 8.5B active and reports 70.2 on Terminal-Bench 2.1. DeepSeek V4 Flash 0731 uses roughly 284B total and 13B active, a one-million-token context and reports 82.7 on the same benchmark. Both remain large, but sparse activation improves token economics compared with similarly capable dense models.

• Laguna has GGUF and MLX builds, although useful quantizations still occupy tens of gigabytes.
• DeepSeek Flash is a strong candidate for code-agent servers and compatible APIs.
• Repeat your own issues and tests with the same harness before adopting either model.

Qwen3.8-Max, Muse Glimmer and Nemotron 3.5 Lightning

Qwen3.8-Max brings Qwen's Max class to a 2.4T-total, roughly 95B-active agentic and multimodal model. Muse Glimmer is a dense 30B model designed to bring coding and tool use below 20 GB at 4-bit. Nemotron 3.5 Lightning uses a highly sparse hybrid architecture, around 30B total and 3B active, to maximize throughput rather than absolute intelligence.

• Qwen3.8-Max belongs on clusters; its 27B variant is the local candidate.
• Muse Glimmer targets 24 GB GPUs and always-on agents.
• Nemotron Lightning fits extraction, classification and high-volume agents.
• Always verify the exact repository, license and revision before downloading.

What local.ai's live benchmarks show

The local.ai snapshot reviewed on August 16 adds what vendor benchmarks cannot: runs on real hardware combining aggregate intelligence, task time, energy and quantization size. Its best runs place DeepSeek V4 Flash 0731 around 77% Intelligence with roughly 90-104 GB quantizations; Laguna S 2.1 ranges from about 57% to 72% depending on the setup; Nemotron 3.5 Lightning reaches roughly 66% at around 22 GB. Qwen3.6 27B and 35B-A3B stand out as efficient references even though they are not the central releases in this list.

• Do not compare percentages without opening the run: model revision, quantization, runtime and hardware differ.
• A smaller quantization may win on time and energy while losing absolute quality.
• Intelligence per joule often favors medium models such as Qwen3.6 35B-A3B and Nemotron Lightning.
• local.ai is a live dashboard: preserve the date and configuration when citing a result.

Mac, RTX or DGX Spark: results that change the purchase

The supplied runs expose the capacity-speed tradeoff. A 128 GB M5 Max can load quantized DeepSeek V4 Flash or Laguna at roughly 35-45 W, but some tasks take five to eleven minutes. A 32 GB RTX 5090 runs 27-35B models much faster, often below two minutes, at higher power and with a strict VRAM ceiling. DGX Spark and RTX PRO 6000 make higher-quality quantizations of large models more comfortable.

• Mac Mini M4 Pro 48 GB: Qwen3.6 35B-A3B offers a strong balance at 6-7 minutes and roughly 18-22 W.
• M5 Max 128 GB: choose capacity and energy efficiency over peak speed.
• RTX 5090: choose speed for models that fit fully in 32 GB.
• Buying hardware purely to save API fees rarely works without measuring volume, privacy value and multi-year total cost.

How to choose without chasing benchmarks

Start with available memory, the task and your tolerance for errors. Run 20 to 50 representative cases and record quality, latency, tokens per second, energy, retries and peak memory. A benchmark winner can still lose on your language, tools or output format.

• 8-16 GB memory: choose smaller or distilled models.
• 24 GB VRAM: compare 27-32B models at 4-bit.
• 64-128 GB unified memory: quantized Laguna or medium MoE models become plausible.
• Cluster deployment: Kimi K3, Qwen3.8-Max and full Inkling become realistic.

Quick comparison

Frequently asked questions

Sources

• local.ai · independent real-hardware benchmarks — Live dashboard reviewed August 16, 2026; inspect each run for hardware, runtime and quantization.
• Kimi K3 model card
• Thinking Machines · Introducing Inkling
• Thinking Machines · Inkling-Small
• Poolside · Laguna S 2.1 model card
• DeepSeek V4 Flash 0731 checkpoint