All Open items under AI Models on AIFIRST News, newest first. 26 items.
Deepseek appears to have quietly phased out Deepseek V4 Pro without an official announcement. Community members noticed the model becoming unavailable on the platform, prompting discussion about the company's model lifecycle practices.
A researcher found that Qwen 3.8B will follow GPT-5.5 Pro reasoning prefill tokens and adopt that model's reasoning style. The finding suggests smaller open models can be steered via prefill injection to produce reasoning patterns characteristic of much larger proprietary models.
OpenBMB released MiniCPM5-2B, which scores 15 on the Artificial Analysis Intelligence Index v4.2, the highest of any open-weights model at 4B parameters or below. The model is available on Hugging Face.
Nvidia has announced a definitive agreement to acquire Hugging Face, the dominant platform for open AI model hosting and collaboration, for $12.9 billion. The deal gives Nvidia direct control over the largest community hub for AI models, datasets, and inference tooling. Community reaction focuses on concerns about platform neutrality and what this means for open-source AI distribution.
IFM released K2 Horizon, a new open model family claiming frontier-level performance with a permissive license. The announcement follows the earlier K2 series and comes with community-produced GGUF quantizations already available. K2 Horizon positions itself as an open alternative to closed frontier models for users running local inference.
DeepSeek has released DeepSeek-V4-Flash-Vision-Exp on Hugging Face, adding vision and multimodal input handling to the existing V4 Flash architecture. The experimental release is available for download and local testing. Community members are running early benchmarks on visual reasoning and code generation from screenshots.
The team behind Apodex 1.1, a model family designed for agentic intelligence on complex tasks, held an Ask Me Anything session on r/LocalLLaMA. The thread covers technical questions about model architecture, context handling, training approach, and planned future releases. The team described their focus on scaling agentic reliability rather than benchmark-optimized general performance.
Chinese AI lab Z.ai has confirmed that Ox Alpha, a model that posted strong benchmark results without public attribution, is built on the GLM series. The company says it will open-weight the model. Bloomberg reported the identification based on technical analysis and company confirmation.
Reports surfaced via a post by hardware analyst Ching-Po Hsu indicating Apple is developing an M5-based server product for AI inference workloads. If accurate, an M5 server would give Apple a unified silicon path from personal devices to cloud-scale inference on the same architecture. No official announcement has been made by Apple.
Xiaomi announced the AI Cube, a prototype three-chip AI inference system combining its Xuanjie O3, O100, and D100 processors in a single package with 1.2 TB/s of aggregate memory bandwidth. The system is positioned to compete with Nvidia-based inference appliances and Chinese domestic AI accelerators. Detailed production timelines and pricing have not been announced.
Qwen 3.8 27B has reached 9th position on Code Arena, a benchmark ranking models by performance on real-world developer coding tasks, while Gemma 4 31B sits at 80th despite having more parameters. The gap illustrates how architecture and training choices matter more than raw parameter count for practical coding output. The result positions Qwen 3.8 27B as competitive with much larger frontier models.
Benchmark results show TielCoder, a 22 GB 4-bit quantized 35B Mixture-of-Experts model, matches Claude Opus 4.6 at medium effort on recent real-world coding tasks and surpasses Kimi K2. The model runs efficiently on constrained hardware and is positioned as a local alternative to frontier cloud models for coding workflows. The comparison used live coding problems rather than standard benchmark suites.
A developer hosted Kimi K3, Moonshot AI's 2.8 trillion parameter model, on 8 NVIDIA B300 GPUs via Modal, achieving 92 tokens per second at $190 per million tokens. Cold boot takes around 27 minutes to load the 1.56 TB model at a cost of roughly $56.79 per hour. The experiment shows frontier-scale inference is now accessible but remains expensive outside managed cloud services.
A developer tested Qwen 3.8 27B against Opus 5 on a hard coding task: porting a 39,000-line procedural shooter written in C to a single-file HTML/Three.js web app. Qwen completed the port and produced a working result. The comparison highlights Qwen 3.8 27B's competitive coding ability for large-scale refactoring work.
A team created Atomic Dynamic GGUF quantizations of Qwen 3.8 27B and compared their performance on a voxel island creation task using an RTX 6000. Results showed strong quality retention across quantization levels. The post shares the quantized files and methodology for running Qwen 3.8 27B locally on consumer hardware.
GLM-4.5-Air, a 106B mixture-of-experts model with only 12B active parameters, now supports Multi-Token Prediction in llama.cpp for a meaningful inference speed boost. The update makes it more practical for users who want a large MoE without requiring multiple high-end GPUs. MTP support is enabled via a configuration flag in the updated build.
A developer used Qwen 3.8 27B to reverse-engineer and document firmware for old hardware, a task that Claude Opus 4 had not been able to complete. Qwen 3.8 27B analyzed the firmware binary, generated documentation, and helped write compatibility shims. The post highlights the model's practical utility for embedded systems and legacy software preservation work.
Liquid AI, known for fast Liquid Foundation Model (LFM) architectures and competitive small language models, is preparing a 100B parameter model release. This would be the company's largest model to date, following its track record of strong performance at smaller parameter counts.
A detailed setup guide documents running Qwen3.8-27B in NVFP4 quantization on a single RTX 5090 under vLLM, achieving 77 tokens/second at short context and 64.7 tokens/second at a real 262K token context window. The post provides reproducible configuration parameters for other 5090 owners targeting long-context local inference.
Palantir CEO Alex Karp publicly accused frontier AI labs of building products designed to create addictive dependency rather than genuine utility, drawing comparisons to drug addiction dynamics. The statement comes amid broader industry debate about engagement-driven business models at OpenAI, Anthropic, and other major labs.
GLM-OCR handles complex tables, code-heavy documents, seals and multilingual layouts with a two-stage layout-then-recognition pipeline. At 0.9B parameters it runs locally via Ollama, vLLM or SGLang.
LLaMA-Factory provides a unified web UI + CLI for fine-tuning LLMs and VLMs without writing custom training code. Supports LLaMA, Mistral, Qwen, DeepSeek, Gemma, Phi, Yi, and 90+ others. Positioned as a no-code altern...
A researcher reported that DeepSeek and Moonshot AI were routing customer conversations to Claude's API without disclosing this to users. The claim raises questions about transparency in AI product stacks and undisclosed third-party model use.
DeepSeek released V4-1 Flash, a multimodal Mixture-of-Experts model with 552 billion backbone parameters and support for up to one million tokens of context. The release follows the earlier V4 Pro and continues DeepSeek's rapid model shipping pace.
OpenUI released OUI-1, a model fine-tuned on DiffusionGemma using OpenUI-Lang, a custom domain-specific language for UI generation. The model generates bespoke interface components rather than raw HTML, targeting design-to-code workflows.
Muse-glimmer-30B ranks competitively against much larger frontier models on EQ-Bench Creative Writing and Hemingway-bench. Users running it locally report strong narrative quality relative to its 30 billion parameter size.
AIFIRST 
