Gemma 4 multimodal agents + Claude Fable 5 GA — Dev Signal
Gemma 4 multimodal agents + Claude Fable 5 GA
Share:
Tool of the Week
Gemma 4 12B runs multimodal agents on laptops
Encoder-free architecture projects audio and vision directly into LLM backbone, cutting memory footprint to 16GB VRAM while matching 26B model reasoning performance.
Developers can now deploy agentic multimodal workflows locally without separate vision/audio encoders, reducing latency and infrastructure costs. Native audio support and sub-26B performance unlock edge deployment patterns previously requiring cloud.
Replaces cloud-dependent multimodal inference and larger models for local workflows. Requires 16GB VRAM minimum; supports Ollama, LM Studio, llama.cpp, vLLM, Hugging Face Transformers. Ready now—Apache 2.0 licensed, weights on HuggingFace/Kaggle, official Skills Repository for agentic patterns included.
Get issues like this in your inbox — free, 3x a week.
Quick Signals
Claude Fable 5 launches on AI Gateway today
Set `anthropic/claude-fable-5` in AI SDK to access a Mythos-class model that sustains multi-day autonomous work with adaptive thinking and higher first-shot correctness on complex problems.
Reduces human check-in overhead on long-running tasks like code review and repository investigation. Parallel sub-agent dispatch and adaptive effort settings let you shift resource allocation from supervision to exception handling.
Replaces prior Claude models for multi-step autonomous work. Requires AI SDK update and Anthropic API key. Ready now—30-day retention policy (no ZDR) is a hard constraint; blocking classifiers on cybersecurity/biology tasks narrow the surface. Worth testing on bug-finding and performance debugging workflows immediately.
“a notable step up over prior Claude models on long-running, ambiguous, multi-step tasks, executing end-to-end on work that previously required frequent human check-ins”
“The model sustains productive output across multi-day runs and dependably dispatches parallel sub-agents”
“Prompts and completions are retained for 30 days and are not used to train Claude”
“set model to `anthropic/claude-fable-5` in the AI SDK”
Gemini 3.5 Live Translate ships speech-to-speech translation
Streaming speech-to-speech model detects 70+ languages, generates continuous translated audio with <5 second latency via Gemini Live API, handles noise-robust inputs without manual language config.
Data Point
Run gpt-oss evals locally with LM Studio uv
Execute OpenAI's AIME 2025 eval suite against gpt-oss-20b running locally via LM Studio using uv for dependency management, yielding detailed HTML/JSON results with 45.4% accuracy on 240 prompts.
Developers can now benchmark reasoning models offline without API calls, capturing full prompt/response traces for debugging. Local eval iteration replaces cloud-dependent testing workflows.
Replaces manual OpenAI API eval runs with self-hosted benchmarking. Requires LM Studio, Python 3.13, uv, and 4+ hour runtime for full 240-prompt suite. Worth trying now if you need local model introspection; increase context length from default 4096 to avoid mid-run failures.
“uv run for the benchmark. This means I get all of the dependencies installed automatically without having to worry about setting up a virtual environment myself”
“the eval suite needs an OpenAI-compatible API to talk to. LM Studio runs one on port 1234”
“the above command runs 240 prompts and can take several hours”
“score is the most important number - the eval suite assigns a 1 for each correct answer and a 0 for incorrect answers and then displays the average”
“Reached context length of 4096 tokens with model (arch: gpt-oss) that does not currently support mid-generation context overflow”
3 issues a week · Free forever · 4,200+ developers
Eliminates turn-by-turn translation bottleneck for real-time multilingual voice apps. Developers can build dubbing and simultaneous multi-language translation without managing complex media streaming infrastructure—platform partners (Agora, LiveKit, Pipecat) handle that layer.
Replaces previous Google Translate limit of 5 languages and English-only routing. Requires Gemini Live API integration (public preview for developers) or app-level integration via Google Translate SDK. Worth trying now if building voice features—early partners (Grab, CJ ENM) report low latency and quality. Private preview for Google Meet; mobile rollout already live.
“Gemini 3.5 Live Translate, our latest audio model for live speech-to-speech translation”
“automatically detects 70+ languages and generates smooth, natural-sounding translated speech that preserves the speakers' intonation, pacing and pitch”
“stays just a few seconds behind the speaker throughout the session”
“Offering 70+ languages, an improvement from the previous limit of just five languages”
“Enabling conversations across over 2000+ language combinations in one meeting”
Claude Fable 5 reaches general availability on AWS
Mythos-class model now broadly available through Bedrock; replaces Claude 3.5 Sonnet for autonomous reasoning and coding tasks at production scale.
Developers building AI applications gain access to state-of-the-art reasoning without tiered access restrictions. Significant benchmark improvements directly impact code generation, knowledge synthesis, and multi-step reasoning workflows.
Ready now via AWS Bedrock API. Requires evaluating pricing vs. your current Claude tier and benchmarking against existing models for your use case. Worth immediate testing if you're already on Bedrock; migration friction is low.
“Claude Fable 5, the first generally available Mythos-class model”
“Claude Fable 5 is state-of-the-art on nearly all tested benchmarks”
“delivers a step-change in autonomous knowledge work and coding”
Native terminal emulator built in Zig with libghostty core library; replaces iTerm2/Alacritty by unifying speed, features, and platform-native UI without compromise.
Eliminates the speed-vs-features tradeoff that forces developers to choose between performance and capability in their primary development tool. Native platform integration (tabs, splits, dock, input methods) reduces friction compared to Electron-based or cross-platform-compromised alternatives.
Ready to try now: 1.0 release is production-ready after 2 years of private beta with 2,000 testers. Requires macOS or Linux (Windows not yet supported). Worth evaluating as drop-in replacement if you use iTerm2, Alacritty, or Kitty. Long-term value unlocks when libghostty stabilizes post-1.0 for embedded terminals and new tools.
“Ghostty 1.0 will be publicly released in December 2024 as an open-source project under the MIT license”
“Ghostty 1.0 aims to be the best drop-in replacement for your current terminal emulator on macOS and Linux”
“Ghostty has been in private beta testing for nearly two years. At the time of writing, Ghostty has around 2,000 testers across macOS and Linux”
“Ghostty supports more of the xterm escape sequences than any other terminal emulator (besides xterm itself)”
“libghostty is the core, cross-platform library that powers Ghostty. It is available as a Zig and C API”
Pin GitHub Actions to commit SHAs, enforce read-only org defaults, isolate secrets per environment—three concrete controls that block the attack surface exploited in Trivy and LiteLLM.
Supply chain attacks via compromised actions or leaked org-level secrets now have measurable friction. Developers can replicate Astral's controls (zizmor audits, pinact automation, branch/tag protections) to reduce blast radius without sacrificing CI/CD velocity.
Replaces loose action pinning and org-wide secret sharing. Requires coordination across dependency graph to hash-pin indirect actions, org-level policy enforcement in GitHub, and manual review of action binaries for immutability gaps. Worth starting now: zizmor and pinact are open-source; GitHub policies are free. Full rollout is non-trivial but high-ROI.
“GitHub Actions has poor security defaults, and security compromises like those of Ultralytics, tj-actions, and Nx all began with well-trodden weaknesses like pwn requests”
“We forbid many of GitHub's most dangerous and insecure triggers, such as pull_request_target and workflow_run, across our entire GitHub organization”
“We require all actions to be pinned to specific commits (rather than tags or branches, which are mutable)”
“we default to read-only permissions at the organization level, and we additionally start every workflow with permissions: {} and only broaden beyond that on a job-by-job basis”
“we use deployment environments and environment-specific secrets. This allows us to further limit the blast radius of a potential compromise”