AI Developer Digest

Mon, Aug 24, 2026

4 signals that cleared the gate16 min read

The Signal — start here

August 24 is an infrastructure day. The headline is NVIDIA Groq 3 LPX entering full production — announced at Hot Chips 2026 — delivering 3,400 output tokens/second on Gemma 4 31B with a 100k-token context, available via Nebius Token Factory with no API migration. Alongside it, OpenAI quietly deprecated the codex mcp-server CLI subcommand: developers using Codex as an MCP server in their tooling pipelines need to migrate to the Codex app server or the Codex plugin for Claude Code. llama.cpp adds DeepSeek V4 multi-GPU tensor support and a Mamba2 GEMV→GEMM dispatch optimization in today's build cadence. The single most important story for the next 48 hours is not new: OpenAI Assistants API shutdown is in 2 days (August 26) — if you haven't migrated off /v1/assistants, /v1/threads, and /v1/runs, today is the last full working day.

Must-reads today

NVIDIA Groq 3 LPX in full production — 3,400 tok/s decode on Gemma 4 31B (100k ctx); 4x faster than nearest alternative for agentic workloads; available via Nebius Token Factory API, no migration required.

⚠️ OpenAI Codex mcp-server deprecated (August 24) — if you run codex mcp-server, migrate to Codex app server today; no code that calls /v1/assistants survives past August 26 either.

Breaking Changes

●Breaking

OpenAI Codex: `codex mcp-server` Command Deprecated

What changed

The mcp-server subcommand of the OpenAI Codex CLI is now deprecated. Codex v0.149.1 ships today as the first version without it. The Codex app server replaces it; for Claude Code users, the Codex plugin for Claude Code is the recommended integration path.

TL;DR

codex mcp-server is gone in v0.149.1; use codex app-server instead, or integrate via the Codex plugin for Claude Code.

Developer signal

If your workflow uses codex mcp-server to expose Codex as an MCP endpoint to other tools, that subcommand is now deprecated and will be removed in a future version — start migration today. The replacement is codex app-server, which runs the Codex app in server mode and exposes the same API surface. If you use Codex from Claude Code specifically, the Codex plugin for Claude Code is now the official integration path. Update any shell scripts, docker-compose files, or CI configurations that invoke codex mcp-server. The @openai/codex npm package is now at v0.149.1 — pin this version after migrating to confirm the new subcommand works in your environment before upgrading further.

Affects you ifYou run codex mcp-server in any shell script, CI pipeline, Docker container, or toolchain wrapper; you expose Codex as an MCP server to other tools.EffortModerate — requires updating invocation scripts and any downstream tooling that references the mcp-server subcommand.

OpenAI Codex Changelog | Date: August 24, 2026 | Link: https://developers.openai.com/codex/changeloghttps://developers.openai.com/codex/changelog

Model Releases

Nothing cleared the quality gate this period. No confirmed new model releases from Anthropic, OpenAI, Google, Meta, Mistral, xAI, DeepSeek, Qwen, or Zhipu/Z.ai on August 23–24, 2026. GLM-5.3 open weights remain on track for ~August 28 (see Worth Watching).

API & SDK Changes

Nothing new beyond the BREAKING section above. OpenAI API, Anthropic platform, and Google Gemini API changelogs show no new entries dated August 23–24. LiteLLM v1.99.0-rc.1 and Ollama v0.33.0-rc2 (both covered August 23) remain latest.

Research

Nothing cleared the quality gate this period. arXiv cs.AI/cs.CL/cs.LG new submissions searched; several papers surfaced (metacognition ensembles, implicit behavioral alignment) but none from recognized labs with code repos and benchmark numbers confirmed within the August 23–24 window. HuggingFace Papers Daily and Simon Willison's weblog (last post August 22) are outside the 24-hour window.

Tooling

Notable

llama.cpp b10604: DeepSeek V4 Multi-GPU Tensor Support

What changed

Adds DeepSeek V4 tensor support with coarser head split granularity and shared expert delayed allreduce communication for multi-GPU setups. Previously, DeepSeek V4's MoE expert routing did not support the sm tensor parallelism mode or optimized cross-GPU expert communication patterns.

TL;DR

DeepSeek V4 multi-GPU inference via llama.cpp now supports -sm tensor with optimized head splits and shared expert delayed allreduce — PR #26490.

Developer signal

If you run DeepSeek V4 on multiple GPUs with llama.cpp and haven't used -sm tensor mode, this adds it. The shared expert delayed allreduce optimization reduces cross-GPU synchronization latency for the shared expert layers in DeepSeek V4's MoE architecture — relevant if you're running V4 in a multi-GPU server environment (2+ GPUs via PCIe or NVLink). Pull the b10604 binary for your platform from the releases page and re-run your throughput benchmarks; the difference is most visible with large batch sizes on 2–4 GPU setups.

Affects you ifYou run DeepSeek V4 locally or on a multi-GPU server via llama.cpp.EffortQuick — pull the new binary; no config changes required, though you may want to explicitly set -sm tensor to use the new mode.

ggml-org/llama.cpp GitHub Releases | Date: August 24, 2026 (07:18 UTC) | Link: https://github.com/ggml-org/llama.cpp/releases/tag/b10604https://github.com/ggml-org/llama.cpp/releases/tag/b10604

Notable

llama.cpp b10605: Mamba2 GEMV → GEMM Dispatch Optimization

What changed

Mamba2 input/output projection layers are now flattened to dispatch GEMM (general matrix multiplication) instead of GEMV (matrix-vector multiplication). Previously, Mamba2 projections were dispatched as vector operations, leaving significant GPU parallelism on the table.

TL;DR

Mamba2 inference on GPU gets better hardware utilization by dispatching projection layers as GEMM ops instead of GEMV — PR #27513.

Developer signal

If you run Mamba2-architecture models via llama.cpp (e.g., Falcon Mamba, Jamba, or any SSM-hybrid model), this is a meaningful throughput improvement for batch inference — GEMM operations allow the GPU to use more compute units simultaneously vs. GEMV's sequential vector passes. No configuration changes needed; upgrading to b10605 or later applies the optimization automatically. The improvement is most pronounced on larger batch sizes (>1 sequence) and larger GPU tiers. For single-token generation latency, improvement may be minimal.

Affects you ifYou run Mamba2, Jamba, Falcon Mamba, or other SSM-hybrid models via llama.cpp on GPU.EffortQuick — upgrade to b10605 or later.

ggml-org/llama.cpp GitHub Releases | Date: August 24, 2026 (08:59 UTC) | Link: https://github.com/ggml-org/llama.cpp/releases/tag/b10605https://github.com/ggml-org/llama.cpp/releases/tag/b10605

Benchmarks & Leaderboards

No new leaderboard entries or benchmark movements confirmed for August 23–24, 2026. Current state (unchanged from prior digest):

ARC-AGI-2: GPT-5.6 Sol 92.5% (#1), Claude Opus 5 90.4% (#2), GPT-5.5 85% (#3). Human average: 66%.
LMArena Text Arena: Claude Fable 5 ~1525 ELO (#1), OpenAI/Anthropic/Google flagship cluster close behind.

lmarena.ai remains egress-blocked; data corroborated via search.

Trends & Emerging Tech

The Decode Bottleneck Goes Hardware-Native

What's happening

NVIDIA Groq 3 LPX entered full production on August 24 (announced at Hot Chips 2026), delivering 3,400 output tokens/second on Gemma 4 31B with a 100,000-token context in Artificial Analysis benchmarking — described as the fastest result recorded for this model. NVIDIA claims 4x faster responsiveness for agentic workloads vs. the nearest alternative. A rack-scale deployment supports up to 256 LP30 accelerators. Nebius Token Factory is first public cloud to offer it, via the same API developers already use (no migration). NVIDIA positions it as an extension of the Vera Rubin platform targeting the "decode" phase specifically — the token-by-token generation step that determines interactive speed for users.

Why watch this

The decode speed delta is relevant to agentic loops: at 3,400 tok/s for a 31B model, multi-step agent plans that currently take seconds per step could compress to sub-second. Nebius's "same API" claim means you can test it today without refactoring — drop-in throughput gain if your workload is token-generation bound rather than prefill/KV-cache bound. Watch whether other inference clouds (Together, Fireworks, Groq.com) match these specs and whether the pricing per-token changes once competitive supply arrives. For developers building latency-sensitive agents (coding agents, chat, tool-use loops), this hardware tier is now publicly reachable.

NVIDIA Newsroom | Date: August 24, 2026 | Link: https://nvidianews.nvidia.com/news/nvidia-groq-3-lpx-now-in-full-production-with-world-class-speed-for-agentic-ai

Technical Discussions

Nothing cleared the quality bar this period. Simon Willison (last post August 22 — outside 24h window), HuggingFace community (egress-blocked), and Hacker News (egress-blocked). No AI-dev HN threads with >200 score confirmed for August 23–24 via search.

Quick Hits

llama.cpp — additional August 24, 2026 builds:

b10606 (Aug 24, ~10:00 UTC) — fix ggml_clamp function and memory allocation utilities. Infrastructure correctness fix; affects any model using clamp operations. https://github.com/ggml-org/llama.cpp/releases/tag/b10606
b10608 (Aug 24, ~11:00 UTC) — fix MP4 file handling when moov atom appears at end; fix SIGPIPE handling on Windows pipes. Relevant if you use llama.cpp with multimodal video inputs on Windows. https://github.com/ggml-org/llama.cpp/releases/tag/b10608
b10610 (Aug 24, 13:18 UTC) — enhanced virtual device naming in CUDA and Metal; Metal device descriptions constructed at initialization. Cleaner device labeling in multi-GPU/multi-device logging. https://github.com/ggml-org/llama.cpp/releases/tag/b10610
b10612 (Aug 24, 13:43 UTC) — disable DOTS3NOTE arch test for WebGPU. Test infrastructure only; no runtime change. https://github.com/ggml-org/llama.cpp/releases/tag/b10612

Worth Watching (Announced, Not Yet Shipped)

Item	ETA	Notes
🚨 2 DAYS: OpenAI Assistants API hard shutdown	August 26, 2026 — TODAY MINUS 2	`/v1/assistants`, `/v1/threads`, `/v1/runs` will return errors. Thread data deleted. Migrate to Responses API + Conversations API. No automated migration tool from OpenAI. Last practical day to migrate: today.
⚠️ 6 DAYS: DALL·E GPT retirement from ChatGPT	August 30, 2026	Download any images you need before this date.
⚠️ 7 DAYS: GPT-5.4/5.4 mini retirement from Codex	August 31, 2026	API-key Codex sessions unaffected. Replace with `gpt-5.6-terra` or `gpt-5.6-luna`.
⚠️ 7 DAYS: Gemini Robotics ER 1.6 Preview shutdown	August 31, 2026	Migrate to `gemini-robotics-er-2-preview`.
⚠️ 8 DAYS: GitHub Copilot model deprecations	September 1, 2026	Check GitHub Copilot settings for affected models.
⚠️ 8 DAYS: OpenAI o3 retired from ChatGPT	August 26, 2026	90-day sunset; same date as Assistants API.
Ollama v0.33.0 stable	Days	Still on rc2 (Aug 23). Claude Desktop integration and KV cache fixes pending stable release.
vLLM v0.28.0 stable	Days	rc2 (Aug 21) is latest. Stable expected imminently.
LiteLLM v1.99.0 stable	Days	rc.1 (Aug 23) is latest. Security-relevant API key hashing fix — wait for stable before promoting.
GLM-5.3 open weights	~Aug 28, 2026	Z.ai (Zhipu) — 744B MoE post-trained, ~40B active. On track for Hugging Face drop after safety review. Watch huggingface.co/zai-org.
Grok 4.7	Late Aug / Early Sep 2026	xAI — "all-around better than 4.6, slightly slower"; 2.1T parameters.
Anthropic TS/Go SDKs at 1.0	~Late Aug 2026	Python SDK hit 1.0 August 20; TypeScript and Go expected to follow.
llama.cpp default server port change: 8080 → 9931	Upcoming (no date set)	Update docker-compose, reverse-proxy configs, and hardcoded port references now.
OpenAI Private Safety Processing (PSP)	September 2026	ZDR-eligible traffic with safety analysis run without OpenAI personnel access.
EU AI Act Article 50 — Watermarking enforcement	December 2, 2026	C2PA + SynthID are the de facto standard. Anthropic text watermark active globally since August 2.

Filtered from 30+ primary sources against a published quality rubric. No press releases, no fluff — only what changes what you build.