AI Technology
June 23, 2026•7 min read
Mistral OCR 4: Structured Document Intelligence
Mistral's June 23 model adds boxes, block types, confidence, 170 languages, self-hosting, and leading OCR scores—with unusually clear benchmark caveats.
Blog
Insights, tutorials, and updates from the ZharfAI team
AI Technology
Engineering, infrastructure, and emerging AI systems
AI Technology
June 16, 2026•7 min read
GLM-5.2: Open Long-Horizon Agents at 1M Context
Z.ai's June 16 MIT release improves terminal, repository, tool-use, and long-running tasks through shared sparse indexing and controllable effort.
AI Technology
June 12, 2026•7 min read
Kimi K2.7 Code: Open Weights for Long-Horizon Coding
Moonshot's June 12 model lifts coding and MCP tool scores over K2.6 while cutting reasoning-token use about 30 percent, with full benchmark footnotes.
AI Technology
June 10, 2026•7 min read
DiffusionGemma: Open Text Generation Beyond Next Token
Google's experimental 26B MoE generates blocks in parallel, exceeds 1,000 tokens per second on H100, and exposes the quality-speed trade-off of text diffusion.
AI Technology
June 9, 2026•7 min read
Claude Fable 5 and Mythos 5: One Model, Two Boundaries
Anthropic's June 9 launch exposes the same frontier model through a guarded general tier and a restricted research tier, changing how benchmarks meet access control.
AI Technology
June 4, 2026•7 min read
Nemotron 3 Ultra: Open 550B Agents at 1M Context
NVIDIA's June 4 model opens weights, data, and recipes for a 550B MoE with 55B active parameters, long context, NVFP4, and high-throughput agents.
AI Technology
June 3, 2026•7 min read
Gemma 4 12B: Native Multimodality in 16GB
Google's June 3 checkpoint brings encoder-free image and audio input, agentic reasoning, and multi-token drafting to a laptop-sized open model.
AI Technology
June 1, 2026•7 min read
MiniMax M3: Open Multimodal Agents with 1M Context
MiniMax combines sparse attention, native image and video input, computer use, and frontier coding in one open-weight model built for long-running work.
AI Technology
May 28, 2026•8 min read
Claude Opus 4.8: Stronger Agents, Coding, and Vision
Anthropic's May 28 flagship improves repository work, long-horizon agents, computer use, and visual reasoning, with a system card that exposes the caveats.
AI Technology
May 20, 2026•8 min read
Command A+: An Open Enterprise MoE on Two H100s
Cohere's May 20 Apache-2.0 release unifies reasoning, vision, tools, retrieval, and 48 languages in a 218B MoE with only 25B active parameters.
AI Technology
May 19, 2026•8 min read
Gemini 3.5 Flash: Frontier Agents at Flash Speed
Google's May 19 model pairs fast inference with strong coding, tool-use, multimodal, and long-horizon scores—but the harness still defines the result.
AI Technology
May 16, 2026•8 min read
Qwen3.7: Max, Plus, and Flash for the Agent Era
Alibaba's Qwen3.7 rollout spans a text flagship, a multimodal agent, and an efficient Flash tier—with one family story and several benchmark caveats.