All posts
Architecture Published Updated 13 min

November 12 agent guide: Gemini 4 Hydrogen, Inkling weights, SGLang

November 12, 2026 closed the window with a broad Gemini 4 worker, an Apache 2.0 open-weight drop, and an inference consortium. Google released Gemini 4 Hydrogen, Thinking Machines published Inkling 2 Small weights, LMSYS launched the SGLang Consortium, and Claude Code 2.1.323 added experimental Code Mode.

Jigar JoshiJigar JoshiAgentic AI Architect and Consultant
November 12, 2026 cover for Gemini 4 Hydrogen and open inference
In this post (8 sections)

November 12, 2026 closes this Bacancy Technology window with three infrastructure-scale moves. Google released Gemini 4 Hydrogen for everyday agentic execution and put Gemini 4 in front of Search users. Thinking Machines opened Inkling 2 Small weights under Apache 2.0. LMSYS launched the SGLang Consortium for open inference. Claude Code 2.1.323 added experimental Code Mode for tool calls through generated JavaScript.

What shipped on November 12

November 12, 2026 primary developments
ReleaseSourceStatusBuilder action
Gemini 4 HydrogenGoogleAPI, Search, Gemini app, AntigravityMigrate 3.8 Flash agent lanes; keep Carbon/Argon for deep work
Inkling 2 Small weightsThinking MachinesApache 2.0 on Hugging FaceBench hosted APIs first; self-host only with 500GB+ VRAM
SGLang ConsortiumLMSYS OrgFounding members liveTrack roadmap notes if you serve open models at scale
Claude Code 2.1.323 Code ModeAnthropicExperimental opt-inPilot in a sandbox; leave org default off until reviewed

Gemini 4 Hydrogen

Hydrogen is Google’s rapid-execution model for coding and agentic workflows, adapted from Gemini 4 Carbon. Google prices it flat across the entire 1M-token context at $1.50 input and $7.50 output per million tokens. Google reports Artificial Analysis Intelligence Index 51 at default high thinking (54 at xhigh), Terminal-Bench 4.1 at 57% versus 18.2% for Gemini 3.8 Flash, and roughly 2.5x faster task completion at high default versus that prior Flash lane.

Orchestrator and multi-model work

Google Antigravity’s Orchestrator can switch between Carbon and Hydrogen in one session. Google reports near-Carbon Terminal-Bench 4.0 success with about 24% less quota and 19% faster completion on average when Orchestrator mixes models. For parallel agents, Carbon can coordinate while Hydrogen powers workers and verifiers. Four thinking levels (low, medium, high default, xhigh) let teams trade latency and cost without changing model ids.

  • Agents’ Last Exam v1.1: 71.6% (model card).
  • AutomationBench-AA: 76.1%.
  • SWE-sweep: 13.5%.
  • Multimodal Knowledge Agent: 53.4%.
  • Safety notes include large cuts in agentic overreach/under-service versus 3.8 Flash and stronger indirect prompt-injection resilience on Gray Swan IPI, per Google.

Inkling 2 Small open weights

Thinking Machines Lab launched Inkling 2 Small on November 5 with closed weights and promised Apache 2.0 after external safety testing. On November 12 the weights landed on Hugging Face with NVFP4 and MXFP8 checkpoints, plus serving recipes for Unsloth, SGLang, vLLM, and TokenSpeed. Architecture: 594B total / 17B active MoE, native image input, 1M context. Hardware bar is high: NVFP4 wants at least 500GB aggregate GPU VRAM. Bacancy Technology recommends hosted endpoints for most teams and self-host only where the cluster already exists.

SGLang Consortium and Claude Code Mode

LMSYS Org launched the SGLang Consortium with NVIDIA, AMD, Intel, Google, Meta, Cloudflare, Modal, and RadixArk at the first SGLang Summit. Members join roadmap meetings; maintainers keep merge control. The same day, Claude Code 2.1.323 added experimental Code Mode so Claude can write JavaScript to call tools, plus MCP SDK v1 compatibility mode because the v1 MCP client is removed, and stricter MCP sign-in confirmation even in bypass and auto modes.

Builder checklist

  1. 01
    Point everyday Gemini agents at Hydrogen
    Keep Carbon and Argon for deep debugging and research. Measure cost per task before making Hydrogen the Antigravity default.
  2. 02
    Add Inkling 2 Small to the open-weight matrix
    Run the same SWE and document tasks used for Kimi K3.1 Pro weights; skip self-host unless VRAM budget clears 500GB.
  3. 03
    Review Claude Code 2.1.323 before fleet upgrade
    Read Code Mode docs, confirm MCP OAuth confirmation behavior, and leave Code Mode off for org defaults until a sandbox pilot passes.

Conclusion

November 12 gives builders a broad Gemini 4 worker, a serious Apache 2.0 open-weight option, and clearer open-inference governance. Read it with November 9’s Fable/Ember split: frontier for hard jobs, Hydrogen and Ember for volume, open weights where clusters and licenses allow.

Need Gemini 4 and open-weight routing reviewed?

Bacancy Technology designs Gemini family routing, open-weight self-host plans, and Claude Code fleet upgrades. Consulting: /consulting/ . Contact: /contact/ .

Sources: Google blog ; thinkingmachines.ai — inkling 2 small ; lmsys.org — 2026 11 12 sglang consortium ; Claude blog .

The weekly take

Agentic AI patterns, delivered Thursdays

What I am shipping, watching, and pruning out of client stacks each week. One email. No fluff.

Shipping an agentic AI project this quarter?
Book a 30-min consult
Frequently asked

Questions readers ask about this post

Share this post
LinkedIn Facebook WhatsApp