Tech Meridian
← ENTITY INDEX
COMPANY · ENTITY #6069
Related event timeline, sources and context from the news index.
Meta’s Superintelligence Lab released Muse Ember, a lightweight, cost-focused AI model designed for agent-style execution tasks (tool calls, coding subtasks, routing, extraction, long-document parsing) and multimodal inputs. Meta cites a cost of $0.10 per million input tokens and positions Ember to handle high-volume work alongside larger models like Muse Wonder.
Thinking Machines says it has released Inkling 2 Small, a mixture-of-experts base model with 594 billion parameters (17 billion active) that accepts text and images and is initially offered via the Tinker platform; the company intends to publish the weights under an Apache 2.0 license after external safety testing. According to Thinking Machines and Artificial Analysis metrics cited by the company, Inkling 2 Small scores 31 on the Artificial Analysis AI Index (an 18-point gain over its predecessor), shows large gains on agentic and coding benchmarks, produces more output tokens and costs more per task, and claims the strongest performance among Western open-weight models.
Tencent has released Hy4, a new foundation model with image understanding and the ability to control software (agents); the model is available through Tencent Cloud, OpenRouter, and Tencent apps while the model weights will be published at an unspecified later date. Tencent says Hy4 is optimized for coding, finance and office work and ranks among the strongest open-weight models, though independent evaluations show mixed results versus Chinese rivals and larger commercial models on hard coding and math tasks.
Mistral released the weights for Mistral Large 4 (Le Chonk), a 1‑trillion‑parameter natively multimodal model, claiming leading open‑weight results on coding and defensive cybersecurity benchmarks and new three-mode reasoning options to trade off cost and performance. The company says the model was trained on European infrastructure, is fully open‑weight, shows gains on FrontierCode (19.7%) and artificial analysis indexes (coding 48%, cyber 51%), and achieves a 96.6% average refusal rate on malicious cyber prompts while supporting self-hosted deployment.
NVIDIA published Nemotron-Personas datasets for France and Belgium — part of its Nemotron data program covering locale-grounded persona datasets for 12 countries — and outlined a workflow for post-training Nemotron models with persona-guided synthetic conversations to improve local accuracy and behavior. The French and Belgian datasets (the latter spanning Dutch, French, German, and English) were created with partners Pleias and KU Leuven to help adapt open models to regional rules, terminology, and cultural nuance.
Artificial Analysis released a revamped Intelligence Index (v5) that uses 60% private test sets, widening the score gap between open and closed models to 12 points (from 3). The update dropped Kimi K3.1 Pro from fourth to ninth and also cut Grok 4.7’s score by 12 points; the AA team revised ambiguous items in its AA-Omniscience knowledge test for the new index.
Grok Voice 3.0, a speech-to-speech model, is being rolled out in the Voice Agent API at the same $0.08/min price as its predecessor. Independent tests (Artificial Analysis) rank it #1 on a Speech-to-Speech Index and report sub‑1s latency and lower cost versus comparable frontier models; it is already deployed across millions of daily conversations in Tesla vehicles and Starlink/SpaceX call centers, with reported high task success and automation rates.
Microsoft released MAI-Transcribe-2-Streaming, a real-time transcription model Microsoft says ranks first for accuracy on Artificial Analysis, supports 60 languages, and delivers first partial results in just over 100 ms; introductory pricing is $0.54 per hour of audio. The company also released MAI-Voice-2.1 and a low-latency variant MAI-Voice-2.1-Flash (about 150 ms) that speak 23 languages, can clone a voice from a few seconds of audio, have built-in safeguards, and are available via Microsoft Foundry, the MAI Playground and OpenRouter (Flash priced at $15 per million characters vs $22 for the standard plan, per Microsoft).
ElevenLabs released Eleven v4, a speech model that better follows direction cues (laughter, whispers, sound effects), maintains voice consistency across long productions, supports over 90 languages and up to 10,000 characters per request, and restores professional voice cloning; an optimized low-latency 'Turbo' variant begins producing audio in about 150 ms. Eleven v4 is available via ElevenAgents, ElevenCreative and the API, with published benchmark and pricing comparisons to rivals and regionally configurable data storage options for enterprise customers.
xAI has made Grok 4.7 available via Amazon Bedrock (served through the bedrock-runtime endpoint with cross-Region inference profiles). The model offers a 500K-token context window, four configurable reasoning-effort levels (low/medium/high/xhigh), improved long-horizon self-verification for coding and knowledge work, new safeguards, and support for Responses, Chat Completions, InvokeModel and Converse APIs.
Epoch AI reports that market prices to achieve fixed scores on five AI benchmarks fell about 47% per quarter (≈13x/year) since 2023, citing examples like OpenAI's o3 vs GPT-5.6; the group warns its sample is limited and measures market price per benchmark score rather than pure algorithmic progress. MIT researchers (Hans Gundlach et al.) using broader model pricing data find 5–10x annual declines in benchmark cost but estimate actual algorithmic efficiency gains of ~3x/year after removing hardware and competition effects, and note some improvements come from higher test‑time compute rather than efficiency.
Xiaomi released the MiMo‑V2.6 lineup, with flagship MiMo‑V2.6‑Pro (1.02T params, 42B active) scoring 46 on Artificial Analysis's Intelligence Index and claiming very low token costs; the company attributes gains to a major expanded reinforcement‑learning run and published an open RL toolkit, training framework, smaller models, and ~7,000 graded training tasks. Separately, Anthropic's threat report says Xiaomi appears in case GTG‑16008, alleging ~400,000 exchanges in March–April 2026 routed MiMo conversations through OpenClaw/OpenCode to Claude to extract training data as part of a broader set of ~190M exchanges across several labs.
xAI says Grok 4.6, a model with a 500K-token context window and four configurable reasoning levels, launched on Amazon Bedrock on August 18, 2026; it is served on both bedrock-mantle and bedrock-runtime and supports the Converse API as well as Chat Completions and Responses. xAI reports Grok 4.6 was trained with supplemental model-generated data and agentic reinforcement-learning tasks, and that it shows improved self-checking, first-pass visual/interactive outputs, and expanded safety testing; xAI published benchmark results including ones from Artificial Analysis.
Allow Google Analytics to measure visits? It stays off until you choose. You can change your choice at any time.