DeepSeek V4.1 Flash Hits 141.3 Tok/s on Two DGX Sparks

The fastest vLLM build yet runs 2.5x the prior best kit, with weights and recipe public.

@sudoingX reports deepseek v4.1 flash flying on 2x DGX Spark, served with vLLM: 141.3 tok/s on structured output, 103.0 on code, 58.3 on prose. He calls it the fastest vLLM build of the model on two Sparks, 2.5x the best prior kit, and ahead of TensorFold v0.6.0 on code and structured output. It serves full context at 262k tokens per request.

The needle test found 3 of 3 at 8k, 32k, 64k, 128k and 256k, the last a 255,709-token prompt, with decode holding between 75 and 82 tok/s. Prefill runs 1,395 to 1,475 tok/s, which he calls his weak side. Each of the 15,360 experts sits at 2 to 5 bits by measured sensitivity, 2.70 on average, and the engine is upstream vLLM main with 22 patches and no fork.

Quality held too: his quantization measures KL divergence of 0.0243 from native, 18% lower than the next best pack, with 95.57% top-1 agreement. It scores closer to native on 7 of 8 domains, and HumanEval+, MBPP+ and IFEval sit within noise. The 196 GB weights and the recipe are public.

From @sudoingX

Agents & Coding

Grok Bot Replaces Search, Automations and Cloud Agents

@emigal says one bot now runs 100% of his searches, automations and Cursor cloud agents.

@emigal reports Grok @Bot as his daily driver: it replaced 100% of his search-like queries on ChatGPT and Claude, and 100% of his Hermes usage and automations. It now runs 100% of his Cursor cloud agents, reviews PRs, and merges with his approval. Its VM links to his Tailscale, reaches his machines over SSH, and can spin up Claude Code or Codex agents in the terminal. He says his main bot can manage other bots and create new ones on the fly.

From @emigal

Almost a Thousand Grok Bot Ideas in One Thread

@poteto's prompt thread collects community use cases, from launch radars to bookmark libraries.

@poteto asked what people got done with Grok Bot this week and collected almost a thousand ideas in one thread. @dr_cintas calls it “BY FAR the best agent right now” and lists 10 copyable routines, starting with a weekday 9am scan for AI launches from the last 24h. @nutlope is turning his X bookmarks into a personal library with the bot, free and open source.

From @poteto

Models & Hardware

Z.ai Ships GLM-Legal for Contract Review and Compliance

Built on GLM-5.3-Flash, the model claims frontier-level scores on mainstream legal benchmarks.

@Zai_org introduces GLM-Legal, built on GLM-5.3-Flash and aimed at real legal workflows: contract review, evidence analysis and compliance. The company says it matches frontier models on mainstream legal benchmarks, and points builders to a tech blog, an API access application, and its AutoClaw tool. @Adi4x4 asks how it compares to Harvey's legal model, and @Rafa_Schwinger asks where the weights are.

From @Zai_org

Sources

Sunday, October 11, 2026 Back to the Archive Tuesday, October 13, 2026