← All digests
AI Developer Digest

Mon, Jun 1, 2026

14 signals that cleared the gate20 min read
The Signal — start here
NVIDIA's Computex 2026 keynote on June 1 was the day's dominant story: Jensen Huang announced Nemotron 3 Ultra (550B-param MoE, 55B active at inference, 48.0 Intelligence Index, 300+ tokens/sec, weights arriving June 4) and Cosmos 3 (first open Physical AI omnimodel, accepts and generates text + images + video + audio + robot actions in a single architecture, available now). For general LLM developers, Nemotron 3 Ultra is the most deployable US open-weights model at this intelligence tier — though Kimi K2.6 (54.0) leads globally among open models, and Opus 4.8 (61.0) leads overall. For robotics and physical-AI developers, Cosmos 3 removes the three-model stack (world model + VLM + policy network) and replaces it with one open, commercially licensed model. Outside Computex, llama.cpp shipped 10+ incremental builds on June 1, including EXAONE 4.5 model support and a spec-decode behavior change that may affect deployments relying on automatic draft-simple activation.
Must-reads today
1
Nemotron 3 Ultra — US open-weights frontier hits 48.0 Intelligence Index at 300+ tokens/sec; weights on HuggingFace/NIM June 4; benchmark context: trails Kimi K2.6 (54.0) and Opus 4.8 (61.0)
2
Cosmos 3 — first open Physical AI omnimodel (32B/8B) with robot action output, available now on GitHub with commercial license; skip unless you're in robotics/AV, but if you are, this is the story of the week
3
Gemini CLI June 18 stop — newly surfaced, not in prior digests: if you have any scripted gemini CLI usage, 17 days to test migration to Antigravity CLI (agy)

Breaking Changes

No breaking changes this period.

Model Releases

2
High

NVIDIA Nemotron 3 Ultra — US Open-Weights Frontier Reaches 48.0 Intelligence Index

What changed
550B-total / 55B-active MoE model replaces Nemotron 3 Super as NVIDIA's flagship open model; Intelligence Index jumps from 36 (Super) to 48.0 (Ultra) while output speed increases to 300+ tokens/second — no prior Nemotron model combined this quality tier with this throughput on a single architecture
TL;DR
550B total / 55B active open-weights MoE, 48.0 Artificial Analysis Intelligence Index, 300+ tokens/sec output, 30% lower inference cost vs. comparable alternatives; weights arriving June 4 on HuggingFace, ModelScope, OpenRouter, and NVIDIA NIM
Developer signal
Weights land June 4 — not yet downloadable as of this digest. Deployment paths: HuggingFace / ModelScope for self-hosted, OpenRouter for managed API, NVIDIA NIM microservice (build.nvidia.com) for drop-in vLLM-style serving on NVIDIA hardware. Critical benchmark context: at 48.0 Intelligence Index, Nemotron 3 Ultra is the strongest US open-weight model, but Kimi K2.6 from Moonshot AI scores 54.0 (13% quality gap) and Opus 4.8 scores 61.0 (27% gap). If you have no US-hosting or export-control requirements, Kimi K2.6 is the stronger open-weights choice on intelligence alone. Nemotron 3 Ultra wins on throughput (300+ tokens/sec vs. slower alternatives at the same quality tier) and on deployment integration with NVIDIA infrastructure. For gpt-oss-120b (33 index) or Gemma 4 31B (39 index) users: Ultra is a substantial direct upgrade. For Opus 4.7/4.8 users evaluating open-weights cost reduction: benchmark the quality delta on your actual task before committing — the 27% Intelligence Index gap is real.
Affects you ifYou are serving open-weight models on NVIDIA GPU infrastructure at scale; you use gpt-oss-120b or Gemma 4 in production and are evaluating upgrades; you have US-hosting or export-control constraints that rule out non-US model weightsEffortQuick (drop-in via NVIDIA NIM or HuggingFace once weights available June 4 — eval first recommended)
High

NVIDIA Cosmos 3 — First Open Physical AI Omnimodel with Native Robot Action Output

What changed
Previous Cosmos series generated video/images of simulated worlds but required a separate policy network to produce robot actions; Cosmos 3 adds native action output — a single forward pass now accepts text, images, video, and ambient audio, and can output any of those modalities plus concrete robot action data (joint angles, gripper positions, trajectory waypoints)
TL;DR
Open Physical AI omnimodel — 32B (Super) and 8B (Nano) variants, #1 across 7+ robotics benchmarks, trained on 20 trillion multimodal tokens including 500k robotics trajectories and 100TB vehicle sensor data; weights and training code available now with commercial license
Developer signal
Weights and training scripts are live at https://github.com/nvidia/cosmos and Hugging Face. Start with Cosmos 3 Nano (8B) for local evaluation on a single high-end workstation GPU; use Super (32B) for production simulation pipelines. The commercial license removes the research-only restriction that blocked Cosmos 1/2 from production deployments. Training data (500k robot trajectories, 455k protein structures, 100TB vehicle sensor data) is also open — use as fine-tuning seed data for domain-specific physical AI. A third Edge variant for real-time on-device inference was announced but not yet released. For developers not building robotic control, simulation, or autonomous vehicle pipelines: this is a Watch item only — the model's differentiated capability is action generation, not text reasoning.

Affects you ifYou are building robot control policies, physics simulators, autonomous vehicle data pipelines, or any system requiring multimodal input → robot action outputEffortModerate (new model architecture with structured action output format; requires understanding joint/trajectory representation schema before integrating into existing robot control stacks)

API & SDK Changes

Nothing new within the 24h scan window. Most recent Anthropic Platform release note: May 29 (Claude Managed Agents on AWS — covered in prior digest). Claude Code CHANGELOG shows no new version released on June 1. anthropic-sdk-python last update: v0.105.2 (May 29 — outside window).

Research

arXiv cs.CL/cs.AI/cs.LG listing pages returned 403 errors at fetch time. HuggingFace Papers Daily returned 403. No papers surfaced via search meeting all quality gate criteria (recognized-lab authorship + associated code + concrete benchmarks + within 24h window) for this period. MiniCPM-V 4.5 (8B multimodal, May 2026) is noted as near-miss — outside scan window.

Tooling

Nothing at full-entry level within the 24h scan window. Five [NOTABLE] llama.cpp builds from June 1 appear in Quick Hits below.

Benchmarks & Leaderboards

1
Notable

Nemotron 3 Ultra Places at 48.0 on Artificial Analysis Intelligence Index — Open-Weights Map Updated

What changed
Nemotron 3 Ultra (48.0) enters the Intelligence Index as the highest-scoring US open-weights model, above gpt-oss-120b (33) and Gemma 4 31B (39); the open-weights global frontier is now: Kimi K2.6 54.0 (China) > Nemotron 3 Ultra 48.0 (US) > Gemma 4 31B 39.0; closed-model leader remains Opus 4.8 at 61.0
TL;DR
US open-weights quality frontier now at 48.0 (Nemotron 3 Ultra); 13% below global open leader Kimi K2.6 (54.0); 27% below Opus 4.8 (61.0)
Developer signal
The index positions matter for model selection decisions under constraints. US-hosted open-weights: Nemotron 3 Ultra is now the benchmark ceiling. Global open-weights: Kimi K2.6 scores 54.0 with no export-control complications for most developers outside defense/government contexts. For teams evaluating "when can we move from closed to open model?": at 48.0 vs. 61.0, you are accepting a 27% quality reduction on the index vs. Opus 4.8. Benchmark this against your task-specific evals, not just the aggregate index. Intelligence Index = aggregate across reasoning, coding, math, instruction following — it will not predict quality gaps on narrow domain tasks.

Affects you ifYou make model selection decisions under cost or sovereignty constraints; you are benchmarking open vs. closed model quality tradeoffsEffortQuick (evaluation only — model weights arrive June 4)

Technical Discussions

Nothing cleared the quality bar this period. LMArena state as of June 1: Grok 4.1 Thinking leads at 1483 Elo (established November 2025, no new model additions confirmed in the 24h window). SWE-bench standings unchanged from prior digest: Claude Mythos Preview 93.9%, GPT-5.5 88.7%, Opus 4.8 88.6%.

Quick Hits

  • llama.cpp b9453 (June 1, 14:56 UTC) — EXAONE 4.5 model support: adds EXAONE 4.5 (LG AI Research) model implementation with vision marker support and GQA for multimodal projection; EXAONE 4.5 multimodal can now be loaded and run locally in llama.cpp. [https://github.com/ggml-org/llama.cpp/releases/tag/b9453]
  • llama.cpp b9452 (June 1, 14:20 UTC) — Vulkan Block-load Q3_K/Q6_K: accelerates quantized matrix operations on Intel Battlemage (BMG) GPUs and NVIDIA via the Vulkan path; performance improvement for INT3/INT6 quantized inference on non-CUDA-primary hardware. [https://github.com/ggml-org/llama.cpp/releases/tag/b9452]
  • llama.cpp b9464 (June 1, 19:57 UTC) — Spec decode behavior change: draft-simple auto-enable feature removed; n_outputs_max fixed. If your llama.cpp deployment relied on automatic speculative decoding activation, you must now enable it explicitly. This is a silent behavior change for any deployment where draft-simple was active without explicit configuration. [https://github.com/ggml-org/llama.cpp/releases/tag/b9464]
  • llama.cpp b9459 (June 1, 18:58 UTC) — Metal GLU kernel templating: Apple Silicon GLU operations now use templated f16/f32 kernels with native type load/store; incremental inference quality/performance improvement for M-series Mac users running quantized models via Metal. [https://github.com/ggml-org/llama.cpp/releases/tag/b9459]
  • llama.cpp b9460 (June 1, 19:23 UTC) — VRAM savings: limits maximum outputs reserved per llama_context sequence; reduces peak VRAM allocation on multi-sequence workloads without changing inference behavior — relevant if you're running at VRAM limits on consumer GPUs. [https://github.com/ggml-org/llama.cpp/releases/tag/b9460]

Worth Watching (Announced, Not Yet Shipped)

10

⚠️⚠️⚠️ Gemini CLI Hard Stop — **June 18 (17 days)** — *First time in this digest*

On June 18, 2026, gemini CLI and Gemini Code Assist IDE extensions stop serving requests for Google AI Pro, Ultra, and free personal users (Enterprise/Standard Gemini Code Assist licenses are unaffected through a later date). The replacement is Antigravity CLI (agy), built in Go, available today. Antigravity CLI retains Agent Skills, Hooks, Subagents, and Extensions (renamed to Antigravity Plugins) but does not have 1:1 feature parity out of the gate. Action this week: Install agy, audit your gemini CLI scripts and CI pipeline steps, and test each against Antigravity. Any script using gemini CLI commands that do not map cleanly to agy equivalents needs a workaround before June 18.

⚠️⚠️⚠️ Gemini API Legacy Schema (Interactions) — Hard Removal **June 8 (7 days)**

(Carried from May 26 digest — 7 days remain, urgent)

The Api-Revision: 2026-05-07 opt-out header stops working June 8. Applications still using response.outputs structure must migrate to response.steps. Action today: search your codebase for response.outputs and Api-Revision: 2026-05-07. 7 days is not much runway.

⚠️⚠️⚠️ Claude Sonnet 4 + Opus 4 — Retirement **June 15 (14 days)**

(Carried from prior digests)

claude-sonnet-4-20250514 and claude-opus-4-20250514 return errors June 15. Migration: Sonnet 4 → claude-sonnet-4-6-20260217; Opus 4 → claude-opus-4-8. Read the migration guide before upgrading to Opus 4.8 — adaptive thinking replaces explicit budget_tokens; setting temperature, top_p, or top_k to non-default values returns a 400 error.

⚠️⚠️ Gemini API Unrestricted Key Deadline — **June 19 (18 days)**

(Carried from prior digests)

All unrestricted Gemini API keys blocked June 19. Restrict via AI Studio → API Keys → "Restrict to Gemini API." Takes 2 minutes; no code changes required.

Google AI for Developers | Link: https://ai.google.dev/gemini-api/docs/api-key

⚠️ GPT-4.5 Retirement from ChatGPT — **June 27 (26 days)** — *Newly surfaced*

GPT-4.5 will be retired from ChatGPT on June 27, 2026. Note: this is specifically the ChatGPT product surface; direct API route retirement is unconfirmed from a primary source as of this digest. If you have any gpt-4.5 model identifiers in code, audit whether they target the API or ChatGPT-based integrations. OpenAI o3 retirement from ChatGPT is also set for August 26, 2026.

OpenAI Platform Changelog | Link: https://platform.openai.com/docs/changelog

Gemini 3.5 Pro — In Internal Use, Expected July 2026

Gemini 3.5 Pro is reportedly in internal use at Google; slated for public release next month. No official announcement yet — only Flash (released May 19, GA) is publicly available. When released, expect significant leaderboard updates. No pricing, model ID, or benchmark numbers disclosed.

Claude Mythos — Public Release Expected "In Coming Weeks"

(Carried from prior digests)

Claude Mythos Preview leads SWE-bench Verified at 93.9% (5.3pp above Opus 4.8). Broad API access pending Anthropic cybersecurity safeguards completion. No GA date, pricing, or model ID disclosed.

Ollama v0.30.0 — Still Pre-Release (rc31 as of May 29)

(Carried from May 15 digest)

v0.30.0 restructures Ollama to use llama.cpp directly as backend, with MLX for Apple Silicon. No stable GA date announced.

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