StepFun 3.7 Flash — A 196B-Parameter MoE, a New Frontier-Class Model That Runs Locally

StepFun has released Step 3.7 Flash, a multimodal Mixture of Experts (MoE) model with 196B total parameters, 11B active parameters, and a built-in 1.8B ViT. It claims throughput of up to 400 TPS and is reportedly capable of running locally on machines with 128GB of RAM.

Its benchmarks show unusually strong results for a flash-class local model: 56.26% on SWE-Bench Pro, 92.82% on DeepSearchQA F1, and 47.2 on HLE w/tools. It also shows substantial improvements over the previous Step 3.5 Flash on agent and tool-use tasks such as Terminal-Bench, Toolathlon, and ClawEval.

One technically interesting trait: the intermediate reasoning (thinking trace) looks almost incoherent, yet the final answers come out "perfect." The "infinite thinking" issue reported in the previous version has also reportedly been fixed.

It is available on Hugging Face in BF16, FP8, NVFP4, and GGUF formats, and a Day-0 support PR for llama.cpp has been submitted. StepFun's decision to contribute directly upstream rather than maintain a fork signals a cooperative stance toward the open source community. For users with 4x3090-class hardware, it could be a practical option.

llama.app Arrives — The Barrier to Local AI Inference Drops Even Lower

Gerganov, the author of llama.cpp, has launched llama.app — an official website, a unified installer, and a single `llama` CLI entry point for llama.cpp, aiming to greatly simplify local AI deployment and third-party agent integration.

The move comes as the local AI ecosystem is rapidly coming together. According to a LangChain survey, as of April 2026 one in three AI teams is running open-weight models — up sharply from one in five nine months earlier. Epoch AI estimates that open-weight models now trail proprietary frontier models by only about four months, and the gap keeps narrowing year over year.

Along similar lines, Ollama has announced "OpenJarvis," envisioning a local-first personal AI aligned with the Stanford/Hazy "Intelligence Per Watt" initiative. NVIDIA has also moved four of its open model families to the Linux Foundation's OpenMDW-1.1 license, working to reduce legal fragmentation across weights, code, documentation, and data.

AI Discovers "Fragnesia," a Critical Linux Kernel Vulnerability

In May 2026, an extremely severe local privilege escalation (LPE) vulnerability dubbed "Fragnesia" (CVE-2026-46300) was discovered in the Linux kernel. Echoing the "Dirty COW" exploit that once shocked the world, it makes it possible to tamper with read-only filesystem caches in memory.

What is most notable is that the vulnerability was identified during autonomous bug hunting by an AI agent. An AI autonomously tracked down and pinpointed a complex regression that would have been difficult to find through conventional fuzzing or human code review.

For system administrators, this is a high-severity vulnerability, and patching is urgent. At the same time, the case demonstrates that AI is becoming a powerful tool that complements human capabilities in security research.

An "Invisible Bug" in Multi-Turn RL — The Token-In-Token-Out Principle

Hugging Face has flagged a subtle but important bug lurking in the training loop of multi-turn RL with tool use.

The core issue: after decoding model output, parsing tool calls, and re-tokenizing the updated conversation, the tokenization shifts, and gradients end up being applied to sequences the model never actually sampled. This "token mismatch" silently undermines training effectiveness.

The proposed fix is the "Token-In, Token-Out" rule — do not re-encode sampled tokens, and maintain a single token buffer across turns. vLLM has also released a native weight-sync API and improvements for asynchronous RL, while fastokens, a Rust-based BPE tokenizer, works to ease the CPU tokenization bottleneck.

Google Managed Agents and Codex on Windows — Vertical Integration of the Agent Stack Accelerates

Google has added "Managed Agents" to the Gemini API. A single API call can provision a sandboxed Linux environment with code execution, web access, and file I/O. On the consumer side, Gemini Spark has launched as a 24/7 personal agent for AI Ultra subscribers in the US.

OpenAI has extended Codex computer use to Windows, with support for remote control from the ChatGPT mobile app. UX improvements such as stable identifiers for background agents and past-chat search are also underway.

The common direction here is a shift from "chatbots" to "managed execution environments with policies and memory." A more vertically integrated agent stack — model + harness + sandbox + UI + remote control + pricing — is becoming the mainstream.

---

**Sources:**

StepFun 3.7 Flash (LocalLLaMA/Hacker News)

llama.app (Gerganov/X)

Epoch AI Research open model lag estimate

Fragnesia CVE-2026-46300 walkthrough (Zenn/you2h)

HuggingFace Token-In-Token-Out RL warning (ClementDelangue/X)

Google Managed Agents / Gemini Spark (Google/X)

OpenAI Codex Windows support (OpenAI/X)

LangChain Deep Agents v0.6 / open model adoption survey