Guide · 2026-08-10

Model of the Month — June 2026 Pick: The Best Local LLM

Last updated 2026-08-10

June 2026 flooded the open-weight scene with new releases. We tested the contenders and name one clear winner for developers running LLMs on their own hardware.

By Mohamed Meguedmi · 8 min read

June 2026 was, by any measure, the busiest launch month the open-weight world has seen. OpenRouter counted four genuinely important open-weight releases from new players across China and the US, all landing within weeks of each other. For readers running models on their own machines rather than renting a closed API, the question is not "what shipped" — it is "what should I actually download this month." This guide answers that with a verdict.

The verdict: June 2026's Model of the Month

After running each candidate through our standard suite, Qwen3-30B-A3B in the Q4_K_M quantization is our Model of the Month. It is not the smartest model released in June — that title belongs to the heavyweight MoEs — but it is the model that most developers should be running, because it delivers the best balance of capability, speed, and hardware cost on the machines people actually own.

The reason comes down to architecture. Qwen3-30B-A3B is a Mixture-of-Experts model: it holds ~30.5B total parameters but activates only about 3.3B per token. That means it reasons like a mid-size dense model while generating tokens at the speed of a tiny one. On a single 24 GB consumer GPU it sustains 90–120 tokens per second at Q4_K_M — fast enough for interactive coding, agents, and local chat without waiting.

Why Qwen3-30B-A3B wins this month

Three things separate it from the pack:

• Quality density. At ~68 MMLU-Pro (community-reported) it beats every dense model that fits in the same VRAM budget. You are not trading much intelligence for the speed.
• Modest footprint. Q4_K_M weights are roughly 18–19 GB. That fits a 24 GB card with room for a 32K+ context window, and it runs — slower — on 16 GB cards with partial offload.
• Permissive licensing. It ships Apache-2.0, so there is no ambiguity for commercial deployment. Verify current terms on the official Qwen3-30B-A3B model card.

The MoE routing is the whole story: you get 30B-class answers while paying 3B-class inference cost. For local hardware, that ratio is the single most important number this month.

The June 2026 contenders

Here is how the month's headline open-weight releases compare. Scores are community-reported and normalized against our methodology; treat them as directional, not absolute.

Model (quant)	Total / Active params	MMLU-Pro*	License	Context
Qwen3-30B-A3B (Q4_K_M)	30.5B / 3.3B	~68	Apache-2.0	128K
gpt-oss-120b (MXFP4)	117B / 5.1B	~73	Apache-2.0	128K
GLM-4.6 (Q4_K_M)	~355B / 32B	~76	MIT	200K
DeepSeek-V3.2 (Q4_K_M)	671B / 37B	~78	MIT	128K

*Community-reported MMLU-Pro, June 2026 snapshots. See our benchmarks hub for the full matrix and raw runs.

The pattern is clear: the two 600B+ and 350B-class MoEs win on raw reasoning, but only the top two rows are realistic for single-GPU or single-workstation local use. gpt-oss-120b is the strongest model that a well-equipped enthusiast can run at home, while GLM-4.6 and DeepSeek-V3.2 effectively require multi-GPU servers or heavy CPU offload with large system RAM.

Hardware and real-world performance

Numbers only matter against the hardware you own. Below are representative throughput figures at Q4_K_M / MXFP4 on common consumer setups. Prices in USD.

Model	Min VRAM (Q4)	Recommended GPU	Approx. tok/s	Approx. GPU cost
Qwen3-30B-A3B	~19 GB	RTX 4090 / 5090 (24–32 GB)	90–120	$1,600–$2,000
gpt-oss-120b	~64 GB	2× 48 GB or 96 GB unified	40–70	$4,000+
GLM-4.6	~200 GB	Multi-GPU server	15–35	$10,000+
DeepSeek-V3.2	~380 GB	8× data-center GPU	10–30	$30,000+

This table is why Qwen3-30B-A3B is the pick for the median reader. It is the only June release that turns a $1,600 graphics card into a genuinely fast, genuinely capable local assistant. Model the electricity and amortized hardware cost for your own usage with our cost calculator before committing to anything larger.

How to run it locally in five minutes

The fastest path is Ollama, which pulls a ready-made Q4_K_M build. These steps assume a 24 GB GPU on Linux, macOS, or Windows.

1. Install Ollama. Download it from the official Qwen3 library page on ollama.com and complete the installer for your OS.
2. Pull the model. Run ollama pull qwen3:30b-a3b. The download is roughly 18–19 GB, so allow time on slower connections.
3. Start chatting. Run ollama run qwen3:30b-a3b and confirm you are getting fast, coherent responses.
4. Tune context. For long agent runs, raise the context window via a Modelfile PARAMETER num_ctx 32768; keep an eye on VRAM headroom.
5. Wire it into tools. Point any OpenAI-compatible client at http://localhost:11434/v1 to use it from your editor or scripts.

Prefer raw GGUF weights and llama.cpp? Grab the quant of your choice from the model card and run with --n-gpu-layers 999 to fully offload. Browse alternative builds and quants in our model catalog.

Who should pick what

One model does not fit every situation. Here is the honest routing.

Your situation	Recommended June 2026 model	Why
Single 24 GB GPU, want the best all-rounder	Qwen3-30B-A3B (Q4_K_M)	Best quality-per-GB and interactive speed. Our Model of the Month.
Enthusiast with 64–96 GB VRAM/unified memory	gpt-oss-120b (MXFP4)	Noticeably stronger reasoning, still self-hostable at home.
Team with a multi-GPU server	GLM-4.6 or DeepSeek-V3.2	Frontier-class open reasoning when hardware is not the constraint.
16 GB GPU or laptop	Qwen3-30B-A3B with partial offload, or a smaller 8B–14B model	Runs, but expect lower tok/s; consider a dense small model for latency.

For a broader shortlist across categories — coding, long-context, and constrained-VRAM builds — see our curated best-of picks. All the underlying scores in this guide are available programmatically through the BestLLMfor public API (licensed CC BY 4.0) and the open-source MCP server, so you can pull the same data straight into your own agents.

Conclusion

June 2026 gave local users more good options than any month before it, but more choice is not the same as more clarity. If you run models on your own hardware, the answer this month is straightforward: download Qwen3-30B-A3B at Q4_K_M. It is the rare release that pairs mid-size intelligence with tiny-model speed and a permissive license, and it does so on hardware people already own. Reach for gpt-oss-120b when you have the memory to spare, and step up to GLM-4.6 or DeepSeek-V3.2 only when a server takes hardware out of the equation. For everyone else, our Model of the Month is the one to run — start with the five-minute install above and check the guides hub for deeper tuning.