AI Intel: Gemma 4 Takes LocalLLaMA + Chinese Open Models Hit 30% + More

Reddit’s AI crowd spent the weekend talking less about one shiny frontier launch and more about a market shift. Google’s Gemma 4 lit up LocalLLaMA, Chinese open models kept gaining real usage share, and the GPT-5.4 versus Claude Opus 4.6 debate kept landing on the same answer: the best model is the one you route to the right step. Open models are not the fallback anymore. They are part of the main stack.

Gemma 4 gave the open-model crowd a real reason to care again

What happened: Google launched Gemma 4 on April 2 under an Apache 2.0 license, and LocalLLaMA reacted like it had been waiting for this one. The lineup includes Effective 2B, Effective 4B, a 26B MoE model, and a 31B dense model. Google says the 31B ranked as the #3 open model on Arena AI’s text leaderboard as of April 1, while the 26B came in at #6. The wider Gemma ecosystem has now crossed 400 million downloads and more than 100,000 community variants.

Why it matters: This is the first open launch in a while that feels aimed at working developers, not just benchmark tourists. Google is selling intelligence per parameter, lower hardware overhead, and agent-friendly behavior. That is a better pitch than raw parameter flexing, and Reddit noticed.

Developer angle: Gemma 4 looks useful for the boring but expensive parts of an AI stack: first-pass coding help, structured output, offline assistants, local evals, and internal tools where you want predictable cost. If you can keep those jobs off a frontier API, your budget gets a lot healthier.

Chinese open models are now a default comparison, not a side conversation

What happened: One of the most quoted stats in this week’s Reddit threads came from an OpenRouter 100-trillion-token study cited by Quartz: Chinese open-source models rose from almost nothing in late 2024 to nearly 30% of weekly usage in some weeks, with roughly 13% average share across the year studied. That matches the discussion on the ground. GLM-5, Kimi K2.5, DeepSeek V3.2, and Qwen keep appearing in real side-by-side tests instead of “interesting if true” rumor threads.

Why it matters: This is not nationalism. It is price-performance. Chinese labs have been shipping fast, improving coding and reasoning quality, and treating long context like table stakes. That pushes the whole market harder than another closed-model teaser ever could.

Developer angle: If your eval matrix still stops at OpenAI, Anthropic, and Google, it is outdated. Most teams need one expensive model for hard steps and one or two cheaper models for everything else. That is where these models fit. And if you want to mix them without rewriting your client for every provider, an OpenAI-compatible endpoint like KissAPI makes that a lot easier.

GPT-5.4 vs Claude Opus 4.6 is mostly a workflow economics story now

What happened: Reddit still cannot agree on a single winner between GPT-5.4 and Claude Opus 4.6. Some developers prefer Opus for code review, writing quality, and long-form reasoning. Others like GPT-5.4 for planning and tool-heavy flows. The clearer difference is pricing: GPT-5.4 is still around $30 per million input tokens and $180 per million output tokens, while Claude Opus 4.6 is closer to $15 input and $75 output. Anthropic also made 1M context generally available for Opus 4.6 and Sonnet 4.6 in March, which keeps Claude strong in long-context work.

Why it matters: Once the top models are close enough, the bill decides the winner. Agents amplify every pricing mistake because one user request often turns into retrieval, planning, tool calls, retries, and review. That is why so many benchmark arguments now end with cost math.

Developer angle: Stop looking for a single forever model. Route cheap models to summarization, extraction, reranking, and first-pass code. Save GPT-5.4 or Opus 4.6 for the last hard step where quality actually pays you back. Also, watch output tokens. That is where a lot of agent teams quietly bleed money.

Netflix’s VOID shows where open AI is heading next

What happened: Netflix quietly open-sourced VOID, short for Video Object and Interaction Deletion. It is built on a 5B CogVideoX base and is designed to remove an object from video while also fixing the consequences of removing it: overlaps, displaced items, secondary motion, the whole mess. The current release targets 384x672 video, up to 197 frames, and the easy path still wants a 40GB-plus GPU. Training reportedly used 8x A100 80GB GPUs.

Why it matters: The point is not that every startup should now self-host video inpainting. The point is that open AI is getting more vertical. The field is moving from “one general model does everything” toward stacks of specialized models that solve one job well.

Developer angle: If you build creator tools, ad workflows, or media products, this is worth tracking. Even if VOID is too heavy for your stack today, the pattern matters: specialist open models are arriving faster, and they are getting good enough to anchor real product features.

Quick Hits

• Claude Haiku 3 retires on April 19. If you still have claude-3-haiku-20240307 pinned anywhere, migrate now.
• Anthropic is cleaning up old 1M beta paths. Sonnet 4 and Sonnet 4.5 lose the old beta route on April 30, while Opus 4.6 and Sonnet 4.6 keep 1M context as standard behavior.
• Edge AI is getting more practical. Gemma 4’s smaller variants make “local model” feel more like phones and lightweight devices, not just a workstation with a monster GPU.