Skip to content
Thursday, 19 November 2026
Tenesys AI News
Subscribe

AI Model Updates

What exactly changed in AI models: context, prices, capabilities. Only values stated by sources — never guessed.

ImportantModelsGoogle Developers · 6 Oct 2026

Google ships EmbeddingGemma 2, an open multimodal embedding model for RAG

ParameterBeforeNow
Base architectureNot specifiedGemma 4-based decoder
ParametersNot specified270M (text/code) up to 740M (full multimodal)
Embedding spaceNot specifiedShared 768-dimensional space across modalities
Context windowNot specified8,192 tokens
Dimension truncationNot specifiedMatryoshka Representation Learning, 768 down to 128 dimensions
Code retrieval benchmark (MTEB Code)EmbeddingGemma 1 baseline+14% score vs EmbeddingGemma 1
LicenseNot specifiedApache 2.0
Read more →
ImportantModelsAWS Machine Learning · 6 Oct 2026

GLM 5.3 (753B MoE) from Z.ai launches on Amazon Bedrock with strong coding and security benchmarks

ParameterBeforeNow
ParametersNot specified753B mixture-of-experts
Coding benchmarksNot specifiedCompetitive on DeepSWE, Terminal Bench 3.0, FrontierSWE; 50% improvement over GLM 5.2 on internal coding benchmark
Cybersecurity benchmark (CyberGym)Not specified84.5 (leading score at release)
Prompt cachingNot specifiedImplicit (automatic) by default, explicit cache breakpoints supported (min. 1,024 tokens per breakpoint)
Cross-Region inferenceNot specifiedUS (us.zai.glm-5.3) and Global (global.zai.glm-5.3) inference profiles
API accessNot specifiedOpenAI-compatible Responses and Chat Completions APIs, plus Bedrock Invoke and Converse APIs
Service tiersNot specifiedFlex, Priority, Standard
Read more →
ImportantVoice AINVIDIA Developer · 1 Oct 2026

NVIDIA Publishes Fine-Tuning Recipe for Dialect-Specific Speech Recognition with Nemotron

ParameterBeforeNow
WER (Najdi+Hijazi test)55.05%29.96%
CER (Najdi+Hijazi test)31.63%12.18%
Full SADA WER58.84%35.61%
FLEURS English WER11.04%10.42%
FLEURS Arabic WER12.67%11.41%
Trainable parameters (top-8 layers unfrozen)Not specified230.4M of 638M total
Language coverage40 language-locales (base model)Same, plus specialized Najdi/Hijazi dialects
Read more →
ModelsHugging Face · 2 Oct 2026

Ai2 Open-Sources AstaBrief, a Fast 8B Model for Scientific Report Generation

ParameterBeforeNow
Base modelNot specifiedQwen3-8B
Report generation time (avg, full pipeline)178.5 seconds (Claude-powered Thinking mode)51.1 seconds (AstaBrief Fast mode)
Speed vs Thinking mode1x (baseline)~3.5x faster
SFT training examplesNot specified47K (from 90K filtered research queries)
DPO training examplesNot specified~6K
Weights/Data availabilityProprietary (Claude 3.5/3.7 Sonnet, o3, o4-mini, GPT-4.1 used for data generation)Open-sourced model weights and training data
Read more →
ImportantAgentsGoogle Developers · 23 Sept 2026

Google's Antigravity SDK Adds Local Offline Model Support with Gemma 4 26B

ParameterBeforeNow
Execution modeCloud-only agentic workflowsLocal/offline execution via LiteRT with Gemma 4 26B A4B, plus hybrid cloud-local orchestration
Hardware requirementNot specifiedRecommended >24GB VRAM or unified memory
Cost in hybrid demoNot specified97.2% of tokens (3,322) run locally/offline; only 95 cloud tokens used for Gemini 3.8 Flash planning
Compatible backendsNot specifiedSupports OpenAI-compatible local servers (Ollama, LM Studio, vLLM) via LocalOpenAIAgentConfig
Read more →