Rio de Janeiro City Hall Releases 397B Open Model "Rio 3.5 Open 397B"

IplanRIO, the IT company of the City of Rio de Janeiro, Brazil, has released **Rio 3.5 Open 397B**, a massive open model post-trained on Qwen 3.5 397B, on Hugging Face. It is available for commercial use under the MIT license and currently ranks high on Hugging Face's trending charts.

The model's biggest distinguishing feature is its integration of a reasoning framework called **SwiReasoning**. Based on Shi et al. (2025), this approach dynamically switches between conventional language-based Chain-of-Thought reasoning (explicit reasoning) and reasoning in latent space (latent reasoning). Through entropy-based confidence estimation, the model explores multiple reasoning paths in latent space during stretches where it is uncertain, then returns to explicit reasoning once confidence recovers. According to the developers, this substantially improves token efficiency while maintaining accuracy.

Against a total of 397B parameters, only about 17B are active (a Mixture-of-Experts configuration). With a context length of 1.01 million tokens, it significantly outperforms its base model Qwen 3.5 397B across multiple benchmarks, scoring 80.2 on SWE-Bench Verified and 90.9 on GPQA Diamond. Notably, it shows gains of +18.3 points on Terminal-Bench 2.1 and +17.0 points on DeepSWE.

It is rare for a public institution to release a frontier-class model under the MIT license, and it is drawing considerable attention in the open-source AI community.

Xiaomi Achieves 1000–3000 tps with MiMo V2.5

Xiaomi has reported reaching an inference speed of **1000–3000 tokens/sec** with its in-house model MiMo V2.5. The news gained traction in Reddit's LocalLLaMA community, and the technical details are laid out in a Xiaomi blog post.

The speedup was reportedly achieved by combining a technique called **DFlash** with a persistent kernel. Models based on DFlash have already been released, and a fully open-source release is planned for the near future.

Speeds exceeding 1000 tps overturn conventional wisdom about LLM inference. For practical text generation, it means response speeds far beyond human reading speed. That said, based on publicly available information alone, the conditions under which these figures were achieved — such as hardware configuration and batch size — cannot be fully verified.

Four Days After Claude Fable 5 Was Pulled, the Impact on Developers in Japan Becomes Visible

The impact of **Claude Fable 5** — released on June 9 and discontinued just three days later on June 12 for being "too powerful" — is now taking concrete shape in Japanese tech blogs.

Multiple articles on Zenn report development experiences that made the most of Fable 5's three days. One developer used Fable 5 to build a programming language called Kei from scratch. Even after the model was discontinued, the code produced during those three days remained, and the developer reflected that losing it "wouldn't have been a problem." Developers' resilience to situations where access to an AI tool is suddenly revoked has also become a topic of discussion.

In another article, a QA engineer compared four models — Claude Sonnet, Opus 4.8, Fable 5, and Codex — on generating test design perspectives. While Fable 5 demonstrated the deepest understanding of the specification, the comparison also demonstrated that every model's prioritization judgments shift as the context grows bloated. It offers interesting practical insight into the interaction between model "strength" and "context length."

As of now, Anthropic has made no official announcement regarding the reasons for Fable 5's suspension or when it might return.

Trends in AI Development Practices in Japan: Context Optimization Becomes a Trend

In Japanese tech blogs, **context optimization** for AI coding tools has become a major theme.

Multiple articles addressed the problem of CLAUDE.md bloat. In response to the challenge that "every rule you add increases the cost of every session," articles introduced cases where cutting the resident context from 635 lines to 239 lines improved both cost and the bug-detection rate, as well as a design philosophy that shifts rules from "always loaded" to "loaded only when needed."

Interest is also growing in **Spec Kit**, a spec-driven development tool from GitHub. It takes an approach that forces AI coding agents to strictly follow a "spec → plan → tasks → implementation" flow, and articles sharing more than six months of real-world usage have started to appear.

Additionally, following GitHub Copilot's transition to an "AI Credits"-based pay-as-you-go model on June 1, articles compiling practical tips for curbing token consumption are also drawing attention.