#LLM

Tencent has released and open-sourced the Hunyuan Hy3 preview, a Mixture-of-Experts (MoE) language model with 295 billion total parameters and 21 billion active parameters, supporting a 256K token context length. This release significantly enriches the open-source LLM ecosystem with a large-scale MoE model focused on complex reasoning and agent tasks, potentially accelerating development in AI-powered coding and scientific applications. The model achieves a 54% reduction in first-token latency for products like CodeBuddy, thanks to deep co-optimization of model architecture and inference framework. It targets enhanced performance in math, science, and code generation.

vLLM v0.24.0 was released with 571 commits from 256 contributors, adding support for the MiniMax-M3 model and extensive performance optimizations for DeepSeek-V4, including a FlashInfer sparse index cache and cluster-cooperative topK kernel. This release significantly expands the model support of vLLM and demonstrates its growing role as a high-performance inference engine for cutting-edge LLMs. The optimization for DeepSeek-V4 improves throughput and latency, benefiting large-scale deployment. The release introduces Model Runner V2 (MRv2) supporting quantized models by default, a new streaming parser engine for tool-call parsing, and integration of DeepEP v2 for expert parallelism. Also, vLLM no longer sets CUDAVISIBLEDEVICES internally, replacing it with a deviceids argument.

vLLM v0.25.0 makes Model Runner V2 the default for all dense models, removes the legacy PagedAttention implementation, and achieves performance parity between the Transformers modeling backend and native vLLM. This release is a major architectural overhaul that simplifies the codebase, eliminates technical debt, and broadens model support, benefiting the LLM inference community with improved performance and compatibility. The release includes 558 commits from 232 contributors, adds new models like LLaVA-OneVision-2 and GLM-5, and introduces a new Streaming Parser Engine for tool-call/reasoning parsing.

Meta publicly launched Muse Spark 1.1 on July 9, 2026, a multimodal AI model designed for agentic coding, with API access available through Meta's developer platform. This release positions Meta as a major competitor to OpenAI and Anthropic in the AI coding assistant space, offering aggressive pricing at $1.25 per million input tokens and potentially disrupting the market by commoditizing coding models. The model is evaluated on Terminal-Bench 2.1, but community analysis highlights that resource limits (6 CPU cores, 8GB RAM) were overridden, which may disqualify results. Pricing is $1.25/$4.5 per million tokens for input/output, with $0.15 for cached input.

Newer Claude models (Opus 4.8 and Sonnet 5) generate malformed tool calls that include extra invented fields, causing tool call rejections in third-party coding harnesses like Pi, while older models did not exhibit this issue. This counterintuitive regression shows that reinforcement learning for specific tool schemas can degrade performance on other tools, posing reliability challenges for developers building agents that rely on consistent tool-use behavior. The malformed calls typically have the correct edit content but include made-up keys in the nested edits[] array, violating the schema. Armin Ronacher hypothesizes that newer models are overtrained on Claude Code's built-in edit tool format.

A researcher released a 10MB LoRA adapter for Qwen3.5-4B that reads internal confidence signals from the model's activations to gate tool-use, reducing hallucination and improving error detection by 0.46 d′ compared to the base model. This approach addresses a key limitation of small language models—their inability to accurately verbalize confidence—by leveraging internal signals, enabling more reliable tool-use and private query handling without requiring larger models. The adapter was trained on 126 hand-authored items and evaluated with confidence intervals; it reduces private query leaks from 22% to 10% and is open-source under Apache-2.0.

Two large memory-management patch sets, developed with LLM assistance by established kernel developers, are being reviewed by the Linux kernel community. This marks a shift in how AI-generated contributions are received, as patches from respected developers are taken seriously, potentially setting a precedent for future LLM-assisted work. One patch set by Rik van Riel introduces 'super page blocks' to reliably allocate 1GB huge pages without the inflexible hugetlbfs reservation system, addressing memory fragmentation challenges.

Huawei has open-sourced OpenPangu-2.0-Flash, a Mixture-of-Experts (MoE) large language model with 92 billion total parameters and 6 billion active parameters, supporting a 512K token context length. The release includes model weights, inference code, and training operations. This open-sourcing provides the community with a high-performance, long-context MoE model from a major tech company, potentially lowering the barrier for researchers and developers to experiment with large-scale MoE architectures. It also signals Huawei's growing involvement in the open-source AI ecosystem. The model uses an MoE architecture where only 6B of the 92B parameters are activated per token, enabling efficient inference. It achieves a 512K token context window, and a larger Pro variant (505B total, 18B active) is slated for release in July.

DeepReinforce released Ornith-1.0, a family of open-weights models (MIT licensed) that achieve state-of-the-art coding performance among open-source models of comparable size, with variants ranging from 9B to 397B parameters. This release pushes forward open-source agentic coding capabilities, enabling developers to run powerful coding assistants locally without vendor lock-in, and the self-scaffolding technique may improve reliability in multi-step tasks. The models are built on top of pretrained Gemma 4 (Apache 2.0) and Qwen 3.5 (Apache 2.0), ensuring license compatibility, and include both dense and Mixture-of-Experts variants. The author tested a 35B GGUF quantized model and reported strong performance on agentic coding tasks.

Fernando Irarrázaval's OpenClaw AI assistant challenge, where 2,000 people attempted to leak secrets via email, ended with zero successful breaches after 6,000 attempts. The underlying model, Anthropic's Opus 4.6, was protected by anti-prompt-injection rules. This experiment provides real-world evidence that frontier models are becoming significantly more robust against prompt injection attacks, a critical AI safety concern. It suggests that security improvements in large language models are translating into practical defenses, though not guaranteeing complete invulnerability. The challenge cost $500 in tokens and triggered a Google account suspension due to excessive inbound emails. Despite 6,000 attempts, no participant managed to leak the secret, but the author warns against deploying production systems where prompt injection could cause irreversible damage.