IBM Granite 4.2 llama.cpp and Claude Code make hybrid local agents shippable

llama.cpp 0.3.0 ships multimodal vision/audio and DeepSeek 4 tensor-split Practical dev impact: Local stacks can now ingest images, audio, and video through `llama-server` / `llama-mtmd-cli` without a cloud projector, and DeepSeek 4 finally splits tensors across GPUs cleanly. The 25 Aug 2026 release adds the dots3-note model with a DSA-ISWA KV cache, WebP via ffmpeg, Pillow-accurate resize, MTP for GLM-4.5-Air, and ggml 0.22.0 (meta-backend tensor split, per-op Metal kernels). Server and web UI gained slot debug knobs and tabbed chat. This is the first llama.cpp version where multimodal is a supported product path rather than a fork.

Claude Code 2.1.246 hardens Auto mode and kills a long tail of session bugs Practical dev impact: Auto mode is now inspectable (new `/permissions` Auto tab) and safer on large transcripts; wildcard Bash allow-rules that silently match extra flags now warn at startup. The 25 Aug 2026 CLI release also timestamps turn completion, truncates pathological diffs, scales safety-check deadlines with prompt size, and fixes a pile of background-session, MCP interrupt, plugin, sandbox, and resume failures that were burning agent hours. Combined with Auto as the default on Pro/Max/Team, the terminal agent is closer to something you can leave on a long SWE loop without a human in the permission loop.

Practical Impact Analysis

Cost and ops shift accordingly. A 30B Granite or a llama.cpp multimodal server is cheaper per SWE-Bench-style loop than a frontier API once you amortize the GPU, and you keep the weights. Mix them: use Claude Code (or Codex) for the high-stakes refactor, Granite or llama.cpp for bulk classification, test generation, and screenshot-to-code. The risk is operational, not capability—thinking-mode tokens, KV-cache shape, and Auto-mode false negatives now live in your runbooks. Teams that still treat “local” as a science-fair project will pay the closed-API tax twice: once in spend, again in data-residency reviews.

Tutorial

# 1. Pull the 8B (or granite4.2:30b if you have the VRAM)
ollama pull granite4.2:8b

# 2. One-shot: thinking on, JSON tool call for a local "run_tests" stub
ollama run granite4.2:8b --think <<'EOF'
You are a coding agent. Think step by step inside <think> tags.
You have one tool:
{"name":"run_tests","description":"Run pytest in cwd","parameters":{"type":"object","properties":{}}}
The repo has a failing test in test_auth.py: test_login_rejects_empty_password.
Propose the one-line pytest invocation you would run, then emit a tool call.
EOF

# 1. Pull the 8B (or granite4.2:30b if you have the VRAM)
ollama pull granite4.2:8b

# 2. One-shot: thinking on, JSON tool call for a local "run_tests" stub
ollama run granite4.2:8b --think <<'EOF'
You are a coding agent. Think step by step inside <think> tags.
You have one tool:
{"name":"run_tests","description":"Run pytest in cwd","parameters":{"type":"object","properties":{}}}
The repo has a failing test in test_auth.py: test_login_rejects_empty_password.
Propose the one-line pytest invocation you would run, then emit a tool call.
EOF

Watch the model emit a `` block, then a structured tool call. Point `OLLAMA_HOST` at a remote box if you want the same binary in CI. Swap `–think` off for latency-sensitive bulk work. Pair with Claude Code Auto for the files you actually merge.

Grok Deep Dive