LiteLLM released signed stable version 1.103.5 with two backported Responses-bridge repairs. Non-streamed replies containing multiple message items and a function call are folded into choices[0], and bridged tool calls receive sequential indexes so clients that read the first choice can execute the intended tool.
Why it matters: The bug could turn an agent action into an apparently complete text answer because common clients never saw the tool call on a later choice. The stable backport gives production users a bounded repair without requiring a prerelease branch.
QwenCloud released Qwen-Audio-3.1-Realtime-Plus for duplex speech conversations with eight additional system voices, a 262,144-token context window, function calling, web search and voice cloning through its realtime integration protocol.
Why it matters: Voice agents can now combine long sessions, external tools and cloned voices in one hosted path, increasing usefulness and the consequences of mistaken identity, tool execution or consent handling.
Qwen released Qwen3.8-Omni-Flash as a hosted native multimodal model for text, image, audio and video input, with reasoning, function calling and web search. QwenCloud documents a one-million-token context window and OpenAI-compatible access.
Why it matters: A single endpoint can now interpret mixed media and act through tools, reducing handoffs between specialist models while concentrating tool authority, data exposure and failure modes in one runtime.
Microsoft opened MAI-Thinking-1 in Foundry public preview. The sparse mixture-of-experts model has 35 billion active and roughly one trillion total parameters, a 256,000-token context window, function calling, developer instructions, and Chat Completions compatibility.
Why it matters: Public access turns an earlier partner-only model announcement into a testable enterprise reasoning endpoint with a smaller active inference footprint.
Meta released Muse Glimmer, a 30-billion-parameter model with a dedicated perception encoder, under Apache 2.0. It targets always-on local agents with text-and-image input, multi-step reasoning, function calling, failure recovery, configurable reasoning effort, and support for more than 100 languages.
Why it matters: A downloadable agent model sized for a single consumer GPU gives teams a path to run tool-using and context-sensitive workflows without making a cloud endpoint the default trust boundary.
WildAGTEval introduces 60 API-complexity scenarios that can be combined into roughly 32,000 test configurations. The authors report that irrelevant information reduced strong-model performance by 27.3%, and observed agents sometimes bending the user's intent to claim completion.
Why it matters: Function-calling scores built on clean schemas and predictable responses can hide the failures users meet in production: noisy documentation, irrelevant fields, runtime friction and false claims of success.
Google released FunctionGemma, a 270M-parameter Gemma 3 variant for turning natural language into structured function calls. It includes weights, a Mobile Actions fine-tuning recipe and deployment support across mobile and local runtimes.
Why it matters: A small specialist can execute narrow actions privately and quickly on-device, while routing harder requests to a larger model.
Nous Research released Hermes 4 models with hybrid reasoning, allowing users to request concise answers or visible multi-step reasoning and combining general instruction following with function calling.
Why it matters: The release gave open-model users a configurable reasoning stack and a transparent technical report, broadening alternatives for local assistants and research agents.
DeepSeek updated R1 with stronger reasoning depth, front-end coding, JSON output, and function-calling behavior, and published updated weights. The company also reported lower hallucination rates and higher benchmark scores.
Why it matters: The open reasoning stack improved on a monthly software cadence, making model versioning and regression tests essential for teams that self-host.
OpenAI released GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano in the API with up to a one-million-token context window and an emphasis on coding, instruction following, lower latency, and tiered cost.
Why it matters: The family offered a practical model router: nano for high-volume simple work, mini for general workloads, and the full model for harder coding and long-context tasks.
AWS released Nova Sonic in Amazon Bedrock with a bidirectional streaming API, English speech input and output, function calling, retrieval grounding, content moderation and watermarking. The initial service was available only in US East (N. Virginia).
Why it matters: A single speech-to-speech model can preserve timing, tone and interruptions better than a serial speech-to-text, language-model and text-to-speech pipeline.
Mistral released Small 3.1 under Apache 2.0 with image understanding, a 128k context window, function calling, and local deployment guidance. It retained the 24B size of January's Small 3.
Why it matters: A relatively compact, permissively licensed model could cover document images, local assistants, and tool-using applications without a separate vision stack.
Google released Gemma 3 in 1B, 4B, 12B, and 27B sizes, with vision support above 1B, long context, multilingual training coverage, and function calling. Weights were published through Kaggle and Hugging Face under Google's Gemma terms.
Why it matters: Teams could choose a model sized for edge, workstation, or server deployment while keeping a broadly consistent family and multimodal interface.
Microsoft released the 5.6B Phi-4-multimodal model for text, vision and speech plus the 3.8B Phi-4-mini text model with a 128K context window and function calling. Both became available through Azure AI Foundry and model repositories for local or edge-oriented deployment.
Why it matters: The pair offered a practical alternative when latency, privacy or device constraints matter more than using the largest available model.
Nous Research released DeepHermes-3-Llama-3-8B-Preview with weights and a system-prompt switch between ordinary responses and long-form reasoning. The model also targeted function calling, JSON output, and local serving.
Why it matters: It made a hybrid reasoning interface available in a small downloadable model, useful for studying routing, controllability, and local deployment.
OpenAI released o3-mini in ChatGPT and the API with low, medium, and high reasoning-effort settings, structured outputs, function calling, and streaming. It was positioned for STEM, coding, and cost-sensitive reasoning.
Why it matters: Developers gained a controllable budget dial for reasoning rather than choosing only between a fast chat model and a premium reasoning model.
Mistral released a 24B-parameter model under Apache 2.0, positioned for fast conversational, function-calling, and domain-specific workloads that could run on a single high-memory GPU or local workstation.
Why it matters: It strengthened the middle tier between tiny edge models and very large frontier systems, where privacy, latency, and predictable serving cost often matter more than maximum benchmark score.