Terminal agents + 975B models: Weekly signals — Dev Signal
Terminal agents + 975B models: Weekly signals
Share:
Tool of the Week
Grok Build ships terminal agent for codebase edits
Rust-based TUI coding agent that edits files, runs shell commands, and searches web interactively or headless via Agent Client Protocol.
Replaces manual context-switching between editor, terminal, and documentation. Enables programmatic codebase manipulation in CI/scripts and embedded workflows without leaving the terminal.
Ready now. Install via one-liner (`curl -fsSL https://x.ai/cli/install.sh | bash`), requires browser auth on first launch. Replaces separate AI chat + manual file edit cycles. Worth evaluating if you're already on macOS/Linux and want agent workflows in existing terminals.
“Grok Build is SpaceXAI's terminal-based AI coding agent”
“understands your codebase, edits files, executes shell commands, searches the web, and manages long-running tasks”
“interactively, headlessly for scripting/CI, or embedded in editors via the Agent Client Protocol (ACP)”
“Prebuilt binaries are published for macOS, Linux, and Windows”
“macOS and Linux are supported build hosts; Windows builds are best-effort and not currently tested from this tree”
Get issues like this in your inbox — free, every weekday.
Quick Signals
Inkling open-weights model reaches 975B parameters
Mixture-of-Experts transformer with 41B active parameters, 1M context window, and native multimodal reasoning available for fine-tuning on Tinker today.
Open-weight release with controllable inference cost via effort scaling eliminates vendor lock-in for multimodal workloads. Tinker integration enables immediate fine-tuning without infrastructure setup.
Replaces closed-weight Claude/GPT for teams needing cost control and customization. Requires GPU infrastructure for inference and fine-tuning; competitive on web dev and agentic tasks but not fastest on individual benchmarks. Worth integrating now if multimodal+efficiency matters more than peak performance.
“975B total parameters, 41B active”
“supports a context window of up to 1M tokens”
“pretrained on 45 trillion tokens of text, images, audio and video”
“Inkling is available for fine-tuning on Tinker today”
“multimodal capabilities, efficient thinking, and availability on Tinker for fine-tuning”
“Sweeping Inkling's effort setting from 0.2 to 0.99 traces its performance against mean generated tokens”
Three open embedding models (8B, 1B BF16, 1B NVFP4) optimize the retrieval-agentic-cost tradeoff; 8B ranks #1 on RTEB, 1B variants reduce error 27–28% over predecessors, NVFP4 achieves 2x throughput on Blackwell.
Retrieval quality directly cuts downstream agent token cost by returning relevant context earlier, reducing repeated searches and reasoning loops. Production deployment now has tiered options: precision (8B), cost-sensitive (1B BF16), and ultra-scale (1B NVFP4) without sacrificing accuracy.
Data Point
GPT-5.6 Sol closes gap on vision benchmarks
Sol reaches 46.2 mAP@50 object detection (up from 13.8 in GPT-5.5), but costs 2.5¢/image and fails on images ≥2000×2000px without resizing workaround.
Vision capabilities now enable practical document workflows, UI automation, and dense-scene detection—but cost and latency tradeoffs mean Gemini 3.5 Flash remains cheaper for high-volume batch processing. Developers building agents must account for 10s latency (Sol) vs 6s (Terra) vs 5s (Luna).
Sol replaces GPT-5.5 for object detection and counting; Terra/Luna offer better cost-latency balance. Requires: resize images >2000×2000px before API calls, use XYXY absolute coordinates (not normalized), budget 0.5–2.5¢ per image. Worth testing now for document extraction and screen understanding; skip for cost-sensitive detection at scale.
“Sol reached 46.2 mAP@50 in our benchmark, while Sol reached 46.2. Terra and Luna followed closely at 44.7 and 43.3”
“Sol becomes less stable on images around 2,000 by 2,000 pixels or larger, especially at lower reasoning effort”
“Sol cost roughly 2.5 cents per image”
“Sol scored 73.0% in our benchmark, up from 64.9% for GPT-5.5”
Replaces Nemotron-3-Embed-1B and generic open embeddings for RAG/agent memory; requires vLLM, NIM microservice, or Hugging Face integration. Ready now: open weights, distillation recipes, and optimized NIM available day-0. Worth trying if you deploy agentic retrieval at scale or need multilingual/code retrieval over 32k windows.
“Nemotron-3-Embed-8B-BF16 ranks #1 on RTEB”
“poor retrieval can cause agents to fetch irrelevant context, re-query, waste token budget, and carry noise into later reasoning steps”
“Nemotron-3-Embed-1B-BF16 brings much of the 8B model's retrieval quality into a smaller deployment footprint”
“reducing error rate by 27% over its 1B predecessor”
“NVFP4 on Blackwell delivers up to 2x higher throughput than BF16 for high-throughput, low-latency retrieval serving”
“retains 99%+ of BF16 retrieval accuracy while reducing memory footprint”
NeMo Automodel enables distributed diffusion training on Diffusers models
NVIDIA and Hugging Face integrated NeMo Automodel to train any Diffusers-format model with FSDP2, tensor, and pipeline parallelism via YAML config—no checkpoint conversion or model rewrites required.
Developers can now fine-tune large diffusion models (FLUX.1-dev 12B, HunyuanVideo 13B) at scale without custom training pipelines or format conversions. Fine-tuned checkpoints round-trip directly back into Diffusers for inference and deployment.
Replaces custom training scripts for Diffusers models. Requires PyTorch DTensor, CUDA dependencies (Docker container provided), and YAML-based configuration. Ready now—includes recipes for FLUX.1-dev, FLUX.2-dev, Wan 2.1/2.2, HunyuanVideo, Qwen-Image. Start with `pip3 install nemo-automodel` or use provided Docker image.
“Pretrained weights from the Hub work out of the box. There's no separate "training format" to convert to, then convert back.”
“Point pretrained_model_name_or_path at any Diffusers model ID on the Hub and start training.”
“Parallelism is a configuration choice, not a code rewrite — switch between FSDP2, tensor parallel, expert parallel, context parallel, and pipeline parallel by declaring configurations, not rewriting models.”
“AutoModel currently supports flow-matching models only.”
“The integration is documented in the Diffusers training guide and is fully open source under Apache 2.0.”
Chat SDK's Slack adapter handles token-by-token streaming, suggested prompts, and feedback buttons—use transcripts instead of channel history for agent conversation context.
Reduces boilerplate for Slack agent UX by handling streaming fallbacks (Post+Edit for GovSlack), prompt pinning, and feedback wiring. Shifts conversation history responsibility to Chat SDK transcripts, forcing a pattern change from native Slack channel history.
Replaces manual Slack event handling for agent threads. Requires Chat SDK v1+ and Slack workspace API access with agent_view entitlements. Ready now for standard Slack; GovSlack users get Post+Edit fallback. Worth adopting if you're already on Chat SDK; greenfield projects should evaluate transcript storage overhead.
“supports the full Slack agent messaging experience, from agent conversations in the Messages tab to suggested prompts, rotating status messages, token-by-token streamed replies, and native feedback buttons”
“Slack threads each user message individually, so channel history only returns the user's side of a DM. Use Chat SDK transcripts to build AI conversation history instead”
“If a workspace doesn't support streaming (e.g., GovSlack), the adapter switches to Post+Edit”
Moonshot's open-source K3 model now routable through Vercel's unified API layer—1M-token window, native multimodal, always-on reasoning mode.
Reduces provider lock-in friction: single SDK call routes to K3 across AI Gateway's retry, failover, and cost-tracking infrastructure. Eliminates context window constraints for long-horizon tasks—code analysis, document reasoning, spatial reasoning workflows no longer need chunking workarounds.
Replaces direct Moonshot API calls with standardized AI SDK routing. Requires: AI SDK v0.x+ and `moonshotai/kimi-k3` model string. Worth trying now if you're already on AI Gateway or evaluating multimodal reasoning for frontend/game dev work. No provider fee markup.
“1M-token context window and native visual understanding, accepting text, image, and video inputs”
“Built for long-horizon software engineering, knowledge work, and deep reasoning”
“Thinking mode is always on”
“AI Gateway reflects provider pricing with no markup and does not charge a platform fee on inference”