As The Geek Learns AI Engineering Digestfrom As The Geek Learns

Monday, August 24, 2026

Claude computer use leaves beta as llama.cpp wires GLM MTP and SenseNova lands on MLX

Anthropic took computer use and browser use out of beta on 19 August 2026. llama.cpp b10603 wired GLM-4.5-Air MTP on 23 August. mlx-community published SenseNova-U1.5 MLX-Swift weights today. Enterprise packaging — Antigravity, AgentCore temporal policies, Unity AI Gateway, Pinecone Nexus — filled the rest of the window.

Window ~2026-07-28 to 2026-08-24. Saturday’s issue already covered the Glimmer / Grok / EU cluster. Primary sources only. Vendor evals flagged.

Breaking AI News

Anthropic computer use leaves beta

19 Aug 2026 Claude Platform notes: computer use is GA as computer_toolset_20260801 (no beta header). Batch actions in one turn; zoom on by default; per-member configs. Same day: browser use (browser_toolset_20260801) via the page accessibility tree. Files API, Agent Skills, and Enterprise Admin user-management also drop beta headers. Upgrading from computer_20251124 changes tool handling and result echoing (toolset_name).

Release notesComputer useBrowser use

llama.cpp b10603: GLM-4.5-Air MTP

Published 2026-08-23 (~2:46 PM ET). PR #26534 wires glm4moe graph_mtp and converter flags --no-mtp / --mtp. Author-reported on 4×3090: ~1.19× with --spec-draft-n-max 1 (74.55 → 88.73 tok/s). Community Strix Halo Vulkan: 21.80 → 27.16 tok/s. No Metal / Mac numbers in the PR or notes. macOS arm64 binary is in the release. vendor

b10603PR #26534

Enterprise

5 stories
5 stories Open section

Local

3 stories

llama.cpp b10603 GLM-4.5-Air MTP

NextN/MTP block was previously unused on existing GLM-4.5 GGUFs. Combined, trunk-only, and MTP-only loaders. Speedups are CUDA/Vulkan, not Metal.

Ollama 0.33.0-rc2

Pre-release 21 Aug: Claude Desktop menu-bar model toggles, Apps view, KV-cache restore-point and Claude Code “tokens left” cache-bust fix.

3 stories Open section

MLX

3 stories

mlx-vlm 0.6.15

18 Aug: batched-row stop fix; tests with mlx 0.32.1. Follow-through for already-covered MLX 0.32.1.

3 stories Open section

Frontier

4 stories

DeepSeek-V4-Flash-Vision-Exp

21 Aug experimental vision model: deepseek-v4-flash-vision-exp. Images billed as input tokens (up to 384 each) at V4-Flash rates.

GPT-5.6 Sol API/credit price cut

21 Aug banner: API and credit pricing of GPT-5.6 Sol dropped by over 20% for the next 3 months. The post does not state the new dollar rates.

4 stories Open section

OSS tools

5 stories

Ray 2.58.0

KV-cache- and token-aware Serve LLM routing finished. Experimental Ray Sandbox; bundled Serve LLM on vLLM 0.26.0.

SGLang 0.5.18

22 Aug: overlapped checkpoint staging; cookbooks for Qwen3.8, Nemotron 3.5 Lightning, DeepSeek-V4-Pro-0813, and several diffusion models.

5 stories Open section

OSS models

4 stories

Ornith-1.5 (MIT)

18 Aug: 397B MoE, 35B-A3B, and 9B dense. Native 262,144 context. Author table vs Claude Opus 4.8 and GLM-5.2. vendor

4 stories Open section

Other

1 story
1 story Open section

Try this on a Mac

2 stories

SenseNova-U1.5 8-bit T2I on MLX-Swift

Weights landed today. Clone sensenova-u1-swift, download the 8-bit card, run sensenova-cli at 1024² / 50 steps. Do not mix the GitHub 3.2 s / 14.8 GB listing with the card’s 7.4 s / 22.9 GB M5 Max figures.

2 stories Open section