The featured models come from new US and Chinese players.
Four open-weight models have crossed into territory where they are powering real agentic pipelines.
New post in our Insights blog about why companies are choosing them in June: https://openrouter.ai/blog/insights/the-open-weight-models-that-matter-june-2026/
Positive users celebrate open-weight models advancing in real agentic pipelines for their strong performance and low cost, while negative users criticize reported speeds as inadequate for coding and question benchmark accuracy.
No Digg Deeper questions have been answered for this story yet.
@OpenRouter What is this BS? Deepseek V4 flash gets 40 on AA but you slapped the score for V4 pro as if its the same model.
@OpenRouter Open weights are becoming practical procurement choices, not ideology.
From the post: DeepSeek V4 Flash for cost, GLM 5.2 for planning, MiniMax M3 for multimodal work, Nemotron when US provenance matters. That is a buying matrix.
@Youssofal_ That row should have said "40" instead of "44 (pro)". Updating the blog post. Thank you!
Four open-weight models have crossed into territory where they are powering real agentic pipelines.
New post in our Insights blog about why companies are choosing them in June: https://openrouter.ai/blog/insights/the-open-weight-models-that-matter-june-2026/
@OpenRouter MiMo v2.5 and MiMo v2.5 Pro are better than DeepSeek V4 Flash/Pro.
@OpenRouter All token / s listed are inadequate for serious coding. Anything below 100 t / s will feel sluggish, basically unusable. For prompt processing you need even more, like above 3000 t / s (yes not a typo. It is 3 thousand tokens per second)
@OpenRouter deepseek-v4-flash is so great and cheap, even my Hermes is struggling to burn my $10 credits nowadays
@OpenRouter Are these models zero data retention?
@OpenRouter Nemotron 3 Ultra is a very good model. I got early access to test it and was pleasantly surprised. Not frontier-level but it’s getting there
@OpenRouter but why no fireworks for minimax m3 even though available throught hem directly
If the gov't & Anthropic keep pushing to gate the strongest models and goes after open-source AI, O.R.'s whole reason for existing gets threatened. Any ideas how to save O.R. or make local AI harder to lock down?
Are you looking at decentralizing hosting, moving infrastructure outside the US, or actively lobbying to protect open-source AI?
The community is ready to back whatever direction actually resists this IMHO possible 1st & 5th amendment violation. What’s the real plan going forward?
@bitcoinmalayali @OpenRouter I meant can i configure openrouter to call these models with zero data retention?
@OpenRouter I feel like ds v4 pro is quite good for agentic work
@rcanand @OpenRouter Depends on the inference provider
@OpenRouter open weights powering agents? cool. still waiting for the demand coordination layer though. this is just supply side flexing
@OpenRouter GLM is excellent, Minimax is good as well
@OpenRouter GLM-5.2 consumed 3,000,000 tokens from a single prompt, which does not seem possible for that task, so I became suspicious. When I ran the same prompt with Opus 4.8, it consumed only 1,200,000 tokens.
@OpenRouter Deepseek flash is lowkey very good
@OpenRouter Construí un sistema agentico que resolvió 50/50 tareas reales de código (Django, scikit-learn, Sphinx, Astropy...) con 100% éxito. Tiempo promedio: 48 segundos. Todo esto corriendo en una Latitude E3540 sin GPU y solo 8GB de RAM. La misma tecnología también optimiza servidores
@OpenRouter DeepSeek V4 Flash has been amazing for me
