AI Daily · Nov 17, 2026: Tencent launches Hy4-Flash: 224B sparse model with 1M context, 50% off

Tencent's Hy4-Flash ships at 224B total/12B active with 1M context and a limited-time 50% discount; OpenAI opens free frontier model access to 200,000 scientists; Qwen4-Max lands on multiple platforms day one.

Pick a date 11/2026: 17 briefings
11/2026
SunMonTueWedThuFriSat
15161718192021
22232425262728
2930

Dotted dates have a briefing; click one to open it.

  • Tencent Hy4-Flash launches with 224B params and 1M context

    Hy4-Flash, Tencent Hunyuan's multimodal model for office and high-frequency tasks, uses 224B total parameters (64B Engram) with 12B active, supports text and image input, a 1M token context window and over 150 tps decoding; it is listed on B.AI with a 50% discount through 2026-11-30 12:00 SGT.

    @BAI_AGI 12 duplicate reports merged
  • OpenAI opens free frontier model access to 200,000 scientists

    OpenAI is expanding free access to its frontier models to up to 200,000 scientists worldwide with training and support; applications open Nov 23 and shared data is not used for training by default.

    @OpenAI
  • Qwen4-Max launches on You.com and Command Code with 1M context

    Qwen4-Max, the first model in Alibaba's Qwen4 series, is available from day one on You.com's Inference Router and in Command Code on all plans. It offers a 1M-token context window, native text, image and video input, and Qwen claims it uses 50% fewer tokens while outperforming Qwen3.8 Max.

    @Alibaba_Qwen 6 duplicate reports merged
  • OpenCode offers $100 of GLM 5.3 usage for $10 to Go subscribers

    OpenCode Go subscribers can get $100 worth of GLM 5.3 usage for $10, part of "Operation GLM Phase 1"; the company says it is served on its own GPUs by its inference team.

    @thdxr 2 duplicate reports merged
  • LoRA adapter boosts Qwen3.8-27B agent scores on SWE-Bench and Terminal-Bench

    A released LoRA adapter for Qwen3.8-27B behind the Pi harness lifts SWE-Bench Pro (Hard 51 subset) pass@1 from 29% to 63% and Terminal-Bench 2.1 from 40% to 56%, with 55-62% fewer tool calls per task.

    @huggingface
  • OpenCode Browse 0.1.0 rewritten in Rust, cuts memory 25x

    OpenCode Browse 0.1.0 is rewritten in Rust with 25x less memory and 23x smaller size, drops the Playwright dependency, and exposes browser actions directly to agents so they can compose them with MCP tools.

    @opencode 3 duplicate reports merged
  • Mistral makes Vibe Pro free for three months in Europe

    Mistral is offering its Vibe coding harness free for the first three months to developers in Europe with code LECHONK, first-come first-served, and claims it beats GLM and DeepSeek at coding.

    @MistralAI 1 duplicate reports merged
  • Ollama details what breaks prefix caching in coding agents

    Ollama reports that a single timestamp in the system prompt can break the cache, changing reasoning effort can wipe the entire cache, and telling an agent its output was truncated makes it re-read missed content. The findings cover Claude Code, Codex, Pi and other harnesses that resend the full conversation each turn.

    @ollama
  • Flint by Recursive goes live and free in Command Code on all plans

    Command Code now offers Flint by Recursive free to all plans and subscribers, with a 262K token context window and 40B total parameters (2.5B active per token).

    @CommandCodeAI 2 duplicate reports merged
  • Trellis.2 reimplemented in CUDA C++ runs 5.3x faster than PyTorch

    A developer ported Trellis.2 fully to CUDA C++ with no Python runtime, reaching 5.3x the speed of the original PyTorch version; moving GLB export to the GPU was the biggest gain, at roughly $7 of compute cost.

    @huggingface