DeepSeek brings app features to its API, including search and voice
The November 13 API update exposes hosted search, document and speech capabilities; it does not announce a new model.
Read the story ↗AiBrainWorX news and models
Free-Range Gaussians reconstructs 3D scenes from sparse views—with separate weight terms
The research project pairs MIT-licensed code with access-controlled weights limited to noncommercial research.
Read the story ↗
Rime’s Coda v2 lets developers specify how words are pronounced
The hosted speech model adds per-request phoneme lexicons and separate streaming endpoints; the announcement dates to October.
Read the story ↗
AWS shows how to evaluate model routing for coding agents
A new walkthrough combines NVIDIA NeMo Switchyard and LiteLLM with Bedrock, while warning that its sample is experimental and incurs cloud costs.
Read the story ↗
Tencent’s Hy ASR 3.0 report examines context-aware speech recognition
A technical report details multilingual transcription and contextual hints; the research paper is new, while the preview predates it.
Read the story ↗
Duo explores real-time audiovisual conversations, with a major license restriction
inclusionAI releases research code and model weights for full-duplex interaction; the model license has territorial limits.
Read the story ↗1 / 6
QwenText
Qwen3-0.6B
An instruction-tuned Qwen text model for conversation, writing, coding and multilingual tasks. It offers thinking and non-thinking modes through its supported chat template.
View model details ↗QwenText
Qwen3-1.7B
An instruction-tuned Qwen text model for conversation, writing, coding and multilingual tasks. It offers thinking and non-thinking modes through its supported chat template.
View model details ↗QwenText
Qwen3-4B
An instruction-tuned Qwen text model for conversation, writing, coding and multilingual tasks. It offers thinking and non-thinking modes through its supported chat template.
View model details ↗QwenText
Qwen3-8B
An instruction-tuned Qwen text model for conversation, writing, coding and multilingual tasks. It offers thinking and non-thinking modes through its supported chat template.
View model details ↗QwenText
Qwen3-14B
An instruction-tuned Qwen text model for conversation, writing, coding and multilingual tasks. It offers thinking and non-thinking modes through its supported chat template.
View model details ↗QwenText
Qwen3-32B
An instruction-tuned Qwen text model for conversation, writing, coding and multilingual tasks. It offers thinking and non-thinking modes through its supported chat template.
View model details ↗QwenText
Qwen3-30B-A3B
An instruction-tuned Qwen text model for conversation, writing, coding and multilingual tasks. It offers thinking and non-thinking modes through its supported chat template.
View model details ↗QwenText
Qwen3-4B-Instruct-2507
An instruction-tuned Qwen text model for conversation, writing, coding and multilingual tasks. This July 2025 Instruct release answers without a separate thinking mode.
View model details ↗QwenText
Qwen3-30B-A3B-Instruct-2507
An instruction-tuned Qwen text model for conversation, writing, coding and multilingual tasks. This July 2025 Instruct release answers without a separate thinking mode.
View model details ↗
HuggingFaceTBText
SmolLM2-135M-Instruct
A compact instruction-tuned language model from Hugging Face’s SmolLM2 family for lightweight text generation and on-device experiments.
View model details ↗HuggingFaceTBText
SmolLM2-360M-Instruct
A compact instruction-tuned language model from Hugging Face’s SmolLM2 family for lightweight text generation and on-device experiments.
View model details ↗HuggingFaceTBText
SmolLM2-1.7B-Instruct
A compact instruction-tuned language model from Hugging Face’s SmolLM2 family for lightweight text generation and on-device experiments.
View model details ↗HuggingFaceTBText
SmolLM3-3B
A 3B instruction model with optional reasoning, six supported languages and long-context training. Hugging Face publishes its training details and data-mixture information.
View model details ↗
mistralaiText
Mistral-7B-Instruct-v0.3
The v0.3 instruction release of Mistral 7B, adding its v3 tokenizer and function-calling support.
View model details ↗mistralaiText
Mistral-Nemo-Instruct-2407
A 12B instruction model jointly developed by Mistral AI and NVIDIA, with multilingual text capabilities and a 128K trained context window.
View model details ↗mistralaiText + images
Mistral-Small-3.1-24B-Instruct-2503
A 24B multimodal instruction model supporting text and image understanding, with a 128K context window.
View model details ↗mistralaiText
Mixtral-8x7B-Instruct-v0.1
An instruction-tuned mixture-of-experts text model that routes each token through selected experts.
View model details ↗
ibm-graniteText
granite-3.3-2b-instruct
An IBM Granite 3.3 instruction model for language tasks including tool use, retrieval-grounded answers, reasoning and multilingual conversations.
View model details ↗ibm-graniteText
granite-3.3-8b-instruct
An IBM Granite 3.3 instruction model for language tasks including tool use, retrieval-grounded answers, reasoning and multilingual conversations.
View model details ↗ibm-graniteText
granite-4.0-h-tiny
IBM’s 7B hybrid instruction model combines Mamba-style and attention layers with sparse experts, long context and instruction following.
View model details ↗
allenaiText
OLMo-2-1124-7B-Instruct
An Allen Institute for AI instruction-tuned OLMo 2 text model, accompanied by open training and research resources.
View model details ↗allenaiText
OLMo-2-0325-32B-Instruct
An Allen Institute for AI instruction-tuned OLMo 2 text model, accompanied by open training and research resources.
View model details ↗
deepseek-aiText
DeepSeek-R1-Distill-Qwen-1.5B
A Qwen-based distilled reasoning model trained using DeepSeek-R1 outputs for mathematical, coding and reasoning tasks.
View model details ↗deepseek-aiText
DeepSeek-R1-Distill-Qwen-7B
A Qwen-based distilled reasoning model trained using DeepSeek-R1 outputs for mathematical, coding and reasoning tasks.
View model details ↗deepseek-aiText
DeepSeek-R1-Distill-Qwen-14B
A Qwen-based distilled reasoning model trained using DeepSeek-R1 outputs for mathematical, coding and reasoning tasks.
View model details ↗deepseek-aiText
DeepSeek-R1-Distill-Qwen-32B
A Qwen-based distilled reasoning model trained using DeepSeek-R1 outputs for mathematical, coding and reasoning tasks.
View model details ↗
microsoftText
Phi-4-mini-instruct
Microsoft’s compact 3.8B instruction model for text reasoning, language understanding and tool-oriented tasks.
View model details ↗microsoftText
Phi-3-mini-4k-instruct
Microsoft’s 3.8B instruction-tuned Phi-3 Mini release with a 4K context window for compact text-generation workloads.
View model details ↗QwenText
Qwen2.5-0.5B-Instruct
An instruction-tuned Qwen2.5 text model for chat, structured answers, coding and multilingual writing.
View model details ↗QwenText
Qwen2.5-1.5B-Instruct
An instruction-tuned Qwen2.5 text model for chat, structured answers, coding and multilingual writing.
View model details ↗QwenText
Qwen2.5-7B-Instruct
An instruction-tuned Qwen2.5 text model for chat, structured answers, coding and multilingual writing.
View model details ↗QwenText
Qwen2.5-14B-Instruct
An instruction-tuned Qwen2.5 text model for chat, structured answers, coding and multilingual writing.
View model details ↗QwenText
Qwen2.5-32B-Instruct
An instruction-tuned Qwen2.5 text model for chat, structured answers, coding and multilingual writing.
View model details ↗DeepSeekText + images
DeepSeek-V4.1-Flash-1017
The October 17 update improves the publisher’s reported coding, reasoning, agent and visual evaluations while reducing output length and task time.
View model details ↗QwenText + images
Qwen3.8-Flash-Next
A multimodal sparse model combining linear and sparse attention, mixture-of-experts routing, n-gram embeddings and multi-token prediction.
View model details ↗QwenText + images
Qwen3.8-27B
A dense multimodal model using a hybrid of Gated DeltaNet and attention layers, with multi-token prediction and selectable reasoning behavior.
View model details ↗
Moonshot AIText + images; video understanding described by publisher
Kimi-K3
A native multimodal mixture-of-experts model for long-running coding, research and knowledge work, using Kimi Delta Attention and Attention Residuals.
View model details ↗
NVIDIAVideo + robot actions
Cosmos3-H-Dreams
An action-conditioned surgical world model that generates video from robot state and action inputs, rather than serving as a general chat assistant.
View model details ↗
GoogleText
gemma-3-270m-it
A very small instruction-tuned Gemma text model for focused tasks on resource-constrained hardware.
View model details ↗GoogleText
gemma-3-1b-it
Google’s compact instruction-tuned Gemma 3 text model for local and on-device applications.
View model details ↗GoogleText
functiongemma-270m-it
A tiny Gemma model specialized for translating requests into function-call structures.
View model details ↗GoogleText → vectors
embeddinggemma-300m
A compact multilingual embedding model for semantic retrieval and similarity, rather than answer generation.
View model details ↗Hugging FaceImages + text
SmolVLM-256M-Instruct
A small vision-language instruction model for image descriptions and questions with limited compute.
View model details ↗Hugging FaceImages + text
SmolVLM-500M-Instruct
A compact vision-language model for image-grounded conversation and lightweight visual experiments.
View model details ↗
hexgradText → audio
Kokoro-82M
A lightweight text-to-speech model with Apache-licensed weights for local speech synthesis.
View model details ↗QwenText → audio
Qwen3-TTS-12Hz-0.6B-Base
The small Qwen3-TTS Base checkpoint supports multilingual speech synthesis from reference voice audio.
View model details ↗QwenText → audio
Qwen3-TTS-12Hz-0.6B-CustomVoice
A small Qwen speech checkpoint for controllable synthesis with supported preset speaker identities.
View model details ↗
KyutaiText → audio
pocket-tts
A small CPU-oriented speech synthesis family for local voice applications.
View model details ↗
KittenMLText → audio
kitten-tts-mini-0.8
A compact Kitten speech model distributed for lightweight local text-to-speech generation.
View model details ↗KittenMLText → audio
kitten-tts-nano-0.8-fp32
A very small Kitten speech model aimed at local synthesis without a large GPU.
View model details ↗
Sentence TransformersText → vectors
all-MiniLM-L6-v2
A small sentence embedding model for English semantic search, clustering and similarity.
View model details ↗
BAAIText → vectors
bge-small-en-v1.5
A small English embedding checkpoint for retrieval-oriented text representations.
View model details ↗
OpenAIAudio → text
whisper-tiny
A tiny multilingual Whisper checkpoint for automatic speech recognition and speech translation.
View model details ↗GoogleImages → labels
mobilenet_v2_1.0_224
A compact MobileNet V2 image classifier pretrained for 224-by-224 image input.
View model details ↗
Wan-AIText → video
Wan 2.1 / small video
A compact text-to-video model for short creative scenes and effects experiments.
View model details ↗Wan-AIImage + audio + optional text → video
Wan 2.2 / speech-to-video
An audio-driven video model combining a reference image, speech and optional scene direction.
View model details ↗QwenCoding
Qwen2.5-Coder-0.5B-Instruct
A very small code-focused instruction model for completion, explanation and short programming tasks.
View model details ↗QwenCoding
Qwen2.5-Coder-1.5B-Instruct
A compact coding assistant candidate for local experiments on modest hardware.
View model details ↗QwenCoding
Qwen2.5-Coder-3B-Instruct
A code-focused model for users who want a middle ground between tiny and larger local coders.
View model details ↗QwenCoding
Qwen2.5-Coder-7B-Instruct
Instruction-following code generation, repair and explanation in a local-sized checkpoint.
View model details ↗QwenCoding
Qwen2.5-Coder-14B-Instruct
A larger dense coder for machines with more memory; select an appropriate quantization.
View model details ↗QwenCoding
Qwen2.5-Coder-32B-Instruct
A dense coding checkpoint for higher-memory workstations and careful offloading experiments.
View model details ↗OpenAISpeech recognition
whisper-base
A compact multilingual speech-to-text checkpoint for local transcription workflows.
View model details ↗OpenAISpeech recognition
whisper-small
A multilingual transcription model offering a different size and accuracy tradeoff from Whisper tiny.
View model details ↗OpenAISpeech recognition
whisper-medium
A larger Whisper transcription checkpoint for users with more compute available.
View model details ↗NVIDIASpeech recognition
parakeet-tdt-0.6b-v2
An English speech recognition checkpoint served through a compatible NeMo or converted ASR runtime.
View model details ↗HuggingFaceTBVision + video
SmolVLM2-256M-Video-Instruct
A small vision-language checkpoint for image and video understanding experiments.
View model details ↗Mistral AIAudio + text
Voxtral-Mini-3B-2507
An audio-language model for transcription and audio understanding with a supported runtime.
View model details ↗Wan-AIVideo generation
Wan2.2-TI2V-5B
A text/image-to-video checkpoint requiring a dedicated video pipeline and careful memory planning.
View model details ↗
SpeakLeash and ACC Cyfronet AGHText
Bielik-11B-v3.1-Instruct
An 11B multilingual text model emphasizing Polish and European languages, with reasoning and tool-call support.
View model details ↗
OpenFold Consortium and AlQuraishi LabBiomolecular structure
OpenFold3-preview
Research-preview biomolecular structure prediction for protein, DNA, RNA and small molecule complexes.
View model details ↗
RunwayRobotics
Praxis-1 DROID
A 27B world-action model checkpoint for DROID-style robot control, predicting video and action chunks.
View model details ↗
inclusionAIMultimodal
Duo-1.0-causal-base
Research checkpoint for full-duplex audiovisual interaction; model terms include territorial restrictions.
View model details ↗
169PiMultimodal
Alpie Seed
9B post-trained Qwen3.5 derivative with text, image and video support in supported implementations.
View model details ↗QwenSpeech recognition
Qwen3-ASR-1.7B
Qwen speech recognition base model; Typd uses a separately licensed fine-tuned adaptation.
View model details ↗
AI at MetaImage to 3D
Free-Range Gaussians
Gated research checkpoint for 3D Gaussian reconstruction from posed image views.
View model details ↗1 / 76
Latest stories
NEWEST FIRST
Anthropic links a public-benefit fund to five percent of its shares
The proposed share commitment raises questions about AI’s wider obligations. Elsewhere, data centers seek a friendlier public image and RTX Spark PCs reach buyers.
Meta revises wording that raised AI-training concerns
Changes to Meta’s language renew attention on how data use is explained, alongside new AI PCs and the limits of genomic interpretation.
Agent incidents and AI-designed viruses test the limits of oversight
For product teams, the important question starts after the headline. MIT Technology Review has placed “Roundtables: A Conversation With the Creator of AI-Designed Viruses” into the current AI conversation. Our role is to separate the…
Indic pretraining reaches a 2.8-trillion-token milestone
The useful signal sits between the announcement and real deployment. Hugging Face has placed “Scaling Indic Pretraining: 2.8 Trillion Tokens and an Open Data Generation Pipeline” into the current AI conversation. Our role is to…
The Nordics build an advanced-computing strategy with NVIDIA
This deserves more than a launch-day reaction. NVIDIA has placed “The Nordics and NVIDIA Show What’s Possible When a Region Commits to Advanced Computing” into the current AI conversation. Our role is to separate the…
A reported office-surveillance case puts chatbot misuse in focus
The useful signal sits between the announcement and real deployment. WIRED AI has placed “An OpenAI Employee Allegedly Used a Chatbot to Help Plant a Recording Device at the Office” into the current AI conversation.…
Meta introduces Muse Wonder to the frontier-model race
The headline is the entry point, not the conclusion. TechCrunch AI has placed “Meta releases its first frontier model, Muse Wonder” into the current AI conversation. Our role is to separate the product or research…
Apple’s HomeView faces the practical test of the connected home
A release becomes meaningful only when it survives contact with actual users. TechCrunch AI has placed “Will the HomeView be Apple’s first smart home success?” into the current AI conversation. Our role is to separate…
Nadella calls for an emergency brake on AI models
For product teams, the important question starts after the headline. TechCrunch AI has placed “Microsoft’s Satya Nadella says AI models need an ‘emergency brake’” into the current AI conversation. Our role is to separate the…
Jev’s reported $7.5 billion valuation spotlights non-text AI
A release becomes meaningful only when it survives contact with actual users. TechCrunch AI has placed “The maker of non-text AI model Jev valued at $7.5B just weeks after launch” into the current AI conversation.…
Frontier agents move deeper into NVIDIA’s developer workflows
The headline is the entry point, not the conclusion. NVIDIA has placed “Into the Omniverse: How Developers Turn Ideas Into Simulations With Frontier AI Agents” into the current AI conversation. Our role is to separate…
Claude Haiku 5.5 reaches AWS as the AI PC story grows
The useful signal sits between the announcement and real deployment. AWS Machine Learning has placed “Introducing Claude Haiku 5.5 on AWS” into the current AI conversation. Our role is to separate the product or research…
Ex-Ramp engineers raise $20M for platform Melius after scrapping their first product
This deserves more than a launch-day reaction. TechCrunch AI has placed “Ex-Ramp engineers raise $20M for platform Melius after scrapping their first product” into the current AI conversation. Our role is to separate the product…
Introducing GLM 5.3 on Amazon Bedrock
For product teams, the important question starts after the headline. AWS Machine Learning has placed “Introducing GLM 5.3 on Amazon Bedrock” into the current AI conversation. Our role is to separate the product or research…