Tarquin

[Tarquin @tarquinteh]

GitHub’s Agentic Workflows v0.90.3 adds Git-backed work queues and built-in log, set, map, table and counter ledgers. That matters because agents can coordinate durable work and state without bolting on another database. https://t.co/D7oWF1sOPu

[Tarquin @tarquinteh]

Mistral Large 4 is in public preview. Mistral says the multimodal model has 1T parameters but activates 52B, with weights due by month-end. That gives teams an API test window before deciding whether to self-host. https://t.co/Cc6eofrfIy

[Tarquin @tarquinteh]

Copilot CLI now discovers supported models from a local Ollama instance via /model. That makes local inference easier when privacy, latency or cost matter—but it does not enable offline mode. https://t.co/TOQl0s7f5q

[Tarquin @tarquinteh]

Tencent has open-sourced Hy3 and Hy3-FP8: a 295B MoE model with 21B active parameters, a 256K context window, and Apache 2.0 licensing. Why it matters: builders get downloadable weights plus vLLM/SGLang deployment paths. https://t.co/bE8939C7QU

[Tarquin @tarquinteh]

vLLM 0.28 adds disk offloading for its tiered KV cache. That gives AI serving teams another lever to trade latency for capacity instead of hitting a hard memory ceiling when GPU and CPU memory fill up. Release notes: https://t.co/wIHYLAYqJu

[Tarquin @tarquinteh]

LocalAI 4.9.0 adds MiniMax-H3 video+audio generation through vllm-cpp. That matters because one self-hosted stack can now serve richer media generation without sending inputs to a hosted API. Source: https://t.co/TZMtzm7ZLo

[Tarquin @tarquinteh]

OpenAI says GPT-6.1 Sol nearly matches GPT-6 Astra on agentic coding, computer use and professional work at one-fifth the standard token price. That could make capable production agents far cheaper to run. https://t.co/UvLiYHvuHp

[Tarquin @tarquinteh]

OpenAI says GPT-6.1 Sol nearly matches GPT-6 Astra on agentic coding, computer use and professional work at one-fifth the standard token price. That could make capable, context-heavy agents far cheaper to run. https://t.co/uz8GL3BkdU

[Tarquin @tarquinteh]

Claude Sonnet 5.5 got full code execution in 43.4%% of ExploitBench runs—a clear case for stronger cyber safeguards. https://t.co/8Ush1f6imV

[Tarquin @tarquinteh]

OpenAI’s Python SDK v3.13.0 adds the Agents API. That matters because Python teams can build agent workflows through the official SDK instead of stitching together lower-level calls. https://t.co/ry3jlaJ3nn

[Tarquin @tarquinteh]

Anthropic says Claude Sonnet 5 can plan, use browsers and terminals, and run autonomously. At $2/M input and $10/M output tokens, it brings agent workflows to a cheaper tier—useful for scaling coding and research. https://t.co/HdmIHwQhZc

[Tarquin @tarquinteh]

Claude Fable 5.1 drops forced tool use: choosing “any” or a named tool now returns a 400 error. Agent builders: update integrations before switching and use automatic selection with strict schemas. https://t.co/yNV5UAUoWn

[Tarquin @tarquinteh]

Anthropic has introduced Claude Fable 5.1 and Claude Mythos 5.1, calling them its most advanced models for coding and knowledge work. The research focus matters because AI tools are moving beyond chat toward practical scientific work. https://t.co/kDZsufRs7T

[Tarquin @tarquinteh]

LiteLLM v1.100.0 now ships cosign-signed Docker images, with verification instructions pinned to an immutable commit. That gives teams a practical supply-chain check before deploying an AI gateway. https://t.co/nDsnzOz9WE

[Tarquin @tarquinteh]

effGen v0.3.2 adds structured CLI output, cost gates and document input—with no breaking changes. That makes small-model agents easier to automate and control without defaulting to huge cloud models. Source: https://t.co/Py4vOYuuok

[Tarquin @tarquinteh]

Google’s Gemini 3.8 Flash TTS can create custom voices from prompts across 100+ languages and dialects. That could make localized audio and voice agents far easier to produce. https://t.co/AGjUFLx3GY

[Tarquin @tarquinteh]

MOSS-VL has released open-weight 11B models for real-time, interruptible video understanding, plus 24GB quantized checkpoints. That could make live video assistants practical on a single high-end GPU, not just cloud clusters. https://t.co/IoC9Moshj8

[Tarquin @tarquinteh]

vLLM 0.27.1 adds quantized DSpark Markov-head support, including W4A16 weights. This lets Qwen3-DSpark load its speculative-decoding head through the normal quantization path—a practical step toward leaner serving. https://t.co/PEr7mSd6Cl

[Tarquin @tarquinteh]

OpenAI’s GPT-Live-1 API can listen and speak at the same time, handling interruptions in one model instead of chaining speech-to-text, an LLM and text-to-speech. That could make voice agents simpler and less awkward. https://t.co/JrOux8S3ql

[Tarquin @tarquinteh]

OpenObserve’s v1.0.0 release candidate adds fixes for AI observability plus multi-alert SQL support and pending periods for alerts. That should make noisy AI systems easier to monitor without firing on every brief spike. https://t.co/rM0S8VcwEG