AI Developer Digest

Tue, Jul 14, 2026

6 signals that cleared the gate36 scanned21 min read

The Signal — start here

Today's digest is dominated by an infrastructure deadline and a new evaluation reality. The gemini-embedding-001 shutdown hits today (July 14) — exactly one year after it launched — and any developer still calling that endpoint will start seeing errors now. The migration requires rebuilding vector indices due to the 768→3072 dimension change. Away from the urgency, llama.cpp continued its punishing build cadence with 17 more builds, adding Hunyuan 3 MTP speculative decoding and fixing a silent DeepSeekV4 sequence removal bug. Ollama v0.32.0 landed with an interactive agent mode and deprecation warnings for six widely-used model families — a meaningful signal that the local inference ecosystem is beginning to clean house on legacy models. The research signal is Long-Horizon-Terminal-Bench, which benchmarks coding agents on 46 long-horizon terminal tasks and finds the best-in-class model still scores only 15% — a stark contrast to the 80%+ figures on SWE-bench Pro.

Must-reads today

gemini-embedding-001 is dead as of today — any calls to this endpoint are now failing; migration to gemini-embedding-2 requires a vector index rebuild (768→3072 dimensions)

llama.cpp b10005 DeepSeekV4 seq_rm fix — if you serve DeepSeek V4-Pro or V4-Flash locally and use sequence removal or KV cache eviction, this was a silent correctness bug; upgrade now

Ollama v0.32.0 deprecation warnings — Llama 3.x, Mistral, Qwen2.5-coder, base DeepSeek-R1, CodeLlama, StarCoder now show deprecation warnings before launching; plan migrations

Breaking Changes

●Breaking

gemini-embedding-001 shutdown — July 14, 2026

What changed

gemini-embedding-001 reached end-of-life today, exactly one year after its July 14, 2025 launch. As of July 14, 2026, all API calls to this model ID return errors. The recommended replacement is gemini-embedding-2.

TL;DR

The Gemini embedding model gemini-embedding-001 is shut down today — calls will now fail; migrate to gemini-embedding-2, which uses 3072 dimensions (up from 768).

Developer signal

If you are calling gemini-embedding-001 anywhere in production, those calls are failing right now. The migration is not drop-in: (1) Rebuild your vector index — gemini-embedding-2 uses 3072 dimensions vs. the 768 dimensions in gemini-embedding-001. Any existing Elasticsearch, Pinecone, Weaviate, Qdrant, or other vector store indices built with 768-dim embeddings are incompatible with 3072-dim output. You must re-embed your corpus and rebuild the index. (2) Update your model ID string — change models/gemini-embedding-001 to models/gemini-embedding-2 (verify the exact model string in the Gemini API documentation, as the deprecations page lists the canonical form). (3) Test similarity scores — embedding dimensions change alters the geometry of your embedding space; retrieval quality will differ and you should re-tune any similarity thresholds you depend on. (4) Check indirect callers — LangChain, LlamaIndex, Firebase Genkit, and n8n integrations that defaulted to gemini-embedding-001 may not have automatically migrated; check your framework version and any plugin that wraps the Gemini embedding API. This shutdown was announced one year ago and the migration window was long, but developer forum posts confirm many teams are being caught by it today.

Affects you ifYou call gemini-embedding-001 directly via the Gemini API or through any framework/library that uses it as a default embedding model (LangChain Google GenAI, LlamaIndex, Firebase Genkit, n8n Google integration).EffortSignificant — requires re-embedding your corpus, rebuilding your vector index, and re-testing retrieval quality.

Google AI for Developers | Date: July 14, 2026 | Link: https://ai.google.dev/gemini-api/docs/deprecationshttps://ai.google.dev/gemini-api/docs/deprecations

Model Releases

No new model releases in the July 13–14 scan window. The active frontier releases remain: GPT-5.6 (July 9), Grok 4.5 (July 9), Claude Fable 5 / Mythos 5 (June 9), Claude Sonnet 5 (June 30), DeepSeek V4 (April 24). All are outside the 24-hour window.

API & SDK Changes

No new API or SDK changes from official lab sources in the July 13–14 scan window beyond what is covered in Breaking Changes above. The active planned changes are in the Worth Watching section.

Research

Medium

Long-Horizon-Terminal-Bench: 46-task benchmark for long-horizon coding agents — best model achieves 15.2%

What changed

A new terminal agent benchmark was submitted that extends prior terminal-task evaluation (Terminal-Bench) with 46 long-horizon tasks using dense intermediate reward grading rather than binary final-outcome scoring. Benchmark code and dataset are publicly available.

TL;DR

Long-Horizon-Terminal-Bench introduces 46 long-horizon terminal tasks across 9 categories with fine-grained subtask grading; the strongest evaluated agent scores 15.2% pass@1 at a 0.95 partial-reward threshold, consuming an average of 9.9M tokens and 85 minutes per task.

Developer signal

If you are evaluating coding agents or building terminal-based agentic systems, this benchmark reveals a capability gap that SWE-bench Pro scores do not: frontier models scoring 70–80% on SWE-bench Pro drop to 15% on tasks requiring sustained, multi-hour terminal interaction with intermediate grading. The nine task categories — including experiment reproduction, software engineering, multimodal analysis, interactive games, and scientific computing — cover scenarios where agents must maintain coherent context across hundreds of tool-call episodes (average: 231 episodes per task). The dense reward signal (scoring intermediate subtasks, not just final outcomes) makes it useful as a training signal target as well as an evaluation harness. The dataset is hosted on Hugging Face; the benchmark harness is at https://github.com/zli12321/LHTB. If you are shipping an agent product and want an honest ceiling estimate for long-horizon tasks, this is the benchmark to run against before claiming production-readiness.

Affects you ifYou are building or evaluating coding agents, software engineering agents, or any system that needs to complete multi-step terminal tasks autonomously over extended timeframes.EffortModerate — requires running the benchmark harness against your agent and setting up the task environment; the harness uses a Terminal-Bench-style setup with reference solutions.

arXiv (2607.08964) / Hugging Face Papers | Date: July 14, 2026 | Link: https://arxiv.org/abs/2607.08964https://arxiv.org/abs/2607.08964 | Code: https://github.com/zli12321/LHTB

Tooling

Notable

llama.cpp b9991–b10007 — 17 builds July 13–14: Hunyuan 3 MTP support, Apple Silicon Q2_0, ARM SME2 kernel, DeepSeekV4 seq fix

What changed

17 builds from b9991 (continuing from b9990, covered July 13) through b10007 (July 14, 19:42 UTC). Developer-relevant highlights from this set:

| Build | UTC | Change |

|-------|-----|--------|

| b10007 | Jul 14, 19:42 | OpenCL: fix dp4a bug for devices where cl_khr_integer_dot_product is unavailable — prevents backend init failure |

| b10005 | Jul 14, 18:57 | DeepSeekV4: fix seq_rm — sequence removal and context update bug corrected |

| b10004 | Jul 14, 18:17 | Vulkan/CPU: f16 SET_ROWS support |

| b10003 | Jul 14, 17:42 | Tokenize tool migrated to common_params_parse |

| b10002 | Jul 14, 15:19 | New tensor contiguity checking functions |

| b10001 | Jul 14, 12:46 | Test utility: fix segfault on no-args invocation |

| b9999 | Jul 14, 11:23 | KleidiAI: add SME2 f32 kernel with dynamic scheduling (ARM) |

| b9996 | Jul 14, 10:41 | Flush log before exit to prevent message loss |

| b9995 | Jul 14, 09:44 | SYCL: set fattn_vec_nthreads=256 for Intel Arc Battlemage detection |

| b9994 | Jul 14, 05:29 | Metal: add Q2_0 support for Apple Silicon |

| b9993 | Jul 13, 23:09 | model: add Hy3 (hy_v3) / Tencent Hunyuan 3 with MTP speculative decoding |

TL;DR

b9993 adds Tencent Hunyuan 3 (open-weight MoE) support with multi-token prediction speculative decoding; b9994 adds Q2_0 quantization to Metal for Apple Silicon; b9999 adds ARM SME2 kernels; b10005 fixes a DeepSeekV4 sequence removal correctness bug.

Developer signal

Four items require action or attention: (1) b10005 — DeepSeekV4 seq_rm bug: If you serve DeepSeek V4-Pro or V4-Flash via llama.cpp and use sequence removal (e.g., KV cache eviction strategies in multi-user serving, context sliding, or the seq_rm API), this was a silent correctness bug affecting sequence copying and context updates. Upgrade to b10005+. (2) b9993 — Hunyuan 3 (hy_v3): Tencent's Hunyuan 3 open-weight MoE architecture is now supported, including MTP (multi-token prediction) speculative decoding. The architecture uses a MoE decoder with per-head Q/K RMSNorm, sigmoid router with expert selection bias, and a permanently-active ungated shared expert. GGUF weights are required; check the Hunyuan model card on Hugging Face for available quantizations. MTP speculative decoding can improve throughput 2–3× for supported models. (3) b9994 — Q2_0 Metal: Q2_0 quantization is now GPU-accelerated on Apple Silicon via Metal. If you are running aggressively quantized models on Apple hardware and hitting CPU offload for Q2_0 layers, this should push those operations to GPU. (4) b9999 — ARM SME2: For ARM CPUs with SME2 support (Cortex-X925 and newer), the new KleidiAI f32 kernel with dynamic scheduling is available. This targets on-device inference on the latest Android SoCs and Arm-based server chips. (5) b10007 — OpenCL dp4a: If you run on OpenCL and hit initialization failure due to cl_khr_integer_dot_product unavailability, upgrade to b10007.

Affects you ifYou serve DeepSeek V4-Pro/Flash locally via llama.cpp with sequence removal enabled; you want to run Tencent Hunyuan 3 GGUF locally; you serve Q2_0-quantized models on Apple Silicon; you run inference on ARM CPUs with SME2 (Cortex-X925+) or Intel Arc Battlemage; you use OpenCL backend on devices without cl_khr_integer_dot_product.EffortQuick — binary or package upgrade; b9993 requires Hunyuan 3 GGUF weights with MTP draft model.

ggml-org/llama.cpp GitHub Releases | Date: July 13–14, 2026 | Link: https://github.com/ggml-org/llama.cpp/releaseshttps://github.com/ggml-org/llama.cpp/releases

Notable

Ollama v0.32.0 — interactive agent experience, deprecation warnings for six major model families

What changed

v0.32.0 stable (July 13) introduces an interactive agent experience activated when running ollama without arguments, and adds deprecation warnings for six commonly-used open model families: CodeLlama, Qwen2.5(-coder), Llama 3.x, Mistral, StarCoder, and base DeepSeek-R1 tags. The Codex App integration was renamed to ChatGPT (ollama launch chatgpt).

TL;DR

Ollama v0.32.0 adds an interactive agent mode (chat/code/web search/delegation via glm-5.2:cloud) when invoked bare, and introduces deprecation warnings before launching CodeLlama, Llama 3.x, Qwen2.5-coder, Mistral, StarCoder, or base DeepSeek-R1.

Developer signal

Two things to act on: (1) Deprecation warnings for commonly-used families: If your application or user workflow launches any of these model families directly from Ollama (ollama run llama3, ollama run mistral, ollama run deepseek-r1, ollama run qwen2.5-coder, ollama run codellama, ollama run starcoder), users will now see a deprecation warning before the model runs. These are not errors — the models still launch — but this is an early signal to plan migrations to current successors (Llama 5, GLM 5.x, DeepSeek V4 family, Qwen3 family). If you are scripting Ollama in non-interactive contexts, these warnings may surface in stderr and surprise integrations that don't handle non-error stderr output. (2) New bare ollama invocation behavior: Running ollama without arguments now launches an interactive agent powered by glm-5.2:cloud rather than showing help text. If you use bare ollama invocation in scripts or aliases as a no-op or help check, this behavior change will surprise you. The integration menu was also streamlined — ollama launch now shows only popular integrations, with the full list still accessible via the same command. ChatGPT integration replaces the "Codex App" label; access via ollama launch chatgpt.

Affects you ifYou use Ollama to serve any of the deprecated model families to end users; you script ollama bare invocations; you integrate Ollama in automation or CI that parses stdout/stderr.EffortQuick (upgrade binary, review invocation scripts for the bare-ollama behavior change).

ollama/ollama GitHub Releases | Date: July 13, 2026 | Link: https://github.com/ollama/ollama/releases/tag/v0.32.0https://github.com/ollama/ollama/releases/tag/v0.32.0

Benchmarks & Leaderboards

No new leaderboard entries or SOTA changes confirmed in the July 13–14 scan window above the quality gate threshold. The LMArena text leaderboard shows Grok-4.1 Thinking at #1 (ELO 1483); no new model additions were confirmed for July 13–14.

Trends & Emerging Tech

Claude for Teachers and the vertical AI product pattern

What's happening

Anthropic launched Claude for Teachers on July 14, offering verified K-12 educators in the US free access to premium Claude capabilities — including Claude Code and Cowork — alongside a curated library of teaching skills, state-standard curriculum integrations (Learning Commons, OpenSciEd, Illustrative Mathematics), and privacy controls (teacher-shared data is not used for model training). The product allows lesson planning from curricula mapped to all 50 state standards, student-differentiation plans with personalized materials, class data analysis, and scheduled recurring tasks like daily exit-ticket review. The timing follows Meta's Muse Spark and Muse Image/Video launches (July 7–9), Google's Classroom AI integrations, and OpenAI's education partnerships — education is the current battleground for AI adoption in the non-developer mainstream.

Why watch this

For developers building in the education vertical, Claude for Teachers represents both an entrant and a template: Anthropic is deploying Claude Code and Cowork into a vertically-specific product with curriculum data integrations, privacy controls, and scheduled automation as table-stakes features. This reveals what Anthropic considers the minimum viable agentic product for a non-developer vertical audience. The pattern to watch: if Claude Code is now being embedded in vertical products (teachers, then likely healthcare providers or enterprise roles), the interface abstraction layer between Claude Code and end-users is thinning. Developers building education-adjacent tools should anticipate competitive pressure from native Anthropic integrations and may find value in MCP connectors that extend Claude for Teachers rather than competing directly.

Anthropic | Date: July 14, 2026 | Link: https://www.anthropic.com/news/claude-for-teachers

Ollama deprecating six model families signals local inference ecosystem is culling legacy weights

What's happening

Ollama v0.32.0 introduced deprecation warnings for CodeLlama, Qwen2.5(-coder), Llama 3.x, Mistral, StarCoder, and base DeepSeek-R1 tags — six families that together represent a large share of active local inference deployments. The warnings are non-blocking (models still run), but the intent is clear: Ollama is guiding users toward current-generation successors. Combined with the Ollama interactive agent mode (which defaults to glm-5.2:cloud rather than a locally-served model), this release represents a notable shift in product posture — from neutral local model runner to opinionated agent runtime with a managed lifecycle.

Why watch this

This is the first time Ollama has deprecated a generation of models en masse rather than individual weights. If the pattern continues, future versions may enforce warnings as errors or remove legacy families from the model library. Developers serving Llama 3.x or Mistral via Ollama to production users should audit their model tags and plan migrations now while the change is still soft. More broadly, the shift toward interactive agent defaults and managed lifecycle signals Ollama is repositioning for a use case beyond "run any GGUF locally" — something closer to a local AI agent platform.

ollama/ollama GitHub Releases | Date: July 13, 2026 | Link: https://github.com/ollama/ollama/releases/tag/v0.32.0

Technical Discussions

Nothing cleared the quality bar for July 13–14. No first-person technical accounts, practitioner posts with primary sources, or community threads with concrete developer data were confirmed published in the window with sufficient primary-source backing.

Quick Hits

Claude for Teachers — Anthropic launched free premium Claude access (including Claude Code + Cowork) for verified K-12 US educators on July 14; includes curriculum integrations for all 50 state standards. https://www.anthropic.com/news/claude-for-teachers
llama.cpp crosses b10000 — The build counter crossed 10,000 today; b10001 and beyond are now the stable tip. This is a minor milestone for the project but useful for version comparisons in changelogs and issue reports. https://github.com/ggml-org/llama.cpp/releases

Worth Watching (Announced, Not Yet Shipped)

Item	ETA	Notes
Claude Fable 5 promo expiry	July 19, 2026	In 5 days: Pro/Max/Team users need usage credits for Fable 5 after 11:59 PM PT; API billing unchanged. Third extension still possible.
Gemini 3.5 Pro	~July 17, 2026	Multiple third-party outlets cite July 17; Google has not confirmed officially. No model card or API endpoint published as of July 14.
Claude Opus 4.7 fast mode removal	July 24, 2026	Requests to `claude-opus-4-7` with `speed: "fast"` will return errors after this date; migrate to Opus 4.8 fast mode.
Claude Opus 4.1 retirement	August 5, 2026	Migrate to Opus 4.8.
Imagen 4 shutdown	August 17, 2026	Migrate to Imagen 4 Ultra or Veo 3.
GPT-Live voice API	"Coming soon"	No pricing, endpoint shape, or date published.

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