Inkling-2-Mini hits 10.04x on KernelBench-Mega locally, beating Grok 4.7 and GPT-6.1
ziqiao_ma · x · 2026-11-23
Inkling-2-Mini, running locally on a single RTX PRO 6000, scored 10.04x (1,650 tok/s) on KernelBench-Mega (02kimilineardecode, isolated Blackwell regrade), beating Grok 4.7 (6.34x), GPT-6.1-Sol (4.96x), DeepSeek-V4-Flash (4.32x), and Claude Sonnet 5 (3.92x), while DeepSeek-V4.1-Flash, Kimi K3.1-Pro, Qwen3.8-Max, and GLM-5.2 failed to pass on that GPU. Claude Opus 5.5 tops the full chart at 35.34x.
Highlights from the 2.9-hour trace (686k output tokens):
- Wrote a single-launch 188-CTA persistent Triton megakernel (1,119 lines) fusing 3 KDA blocks, 1 MLA block, and 4 MoE blocks with global release/acquire atomic barriers (0.75µs sync) and zero inter-op kernel launches
- Its first working fused kernel hit only 3.12x; it deliberately ran a broken no-barrier build to isolate sync overhead, then bisected phases by skipping each one
- Used the MLA absorb trick so the [ctx, 8192] BF16 matrix is never materialized, merging softmax + A + o
More from coding & agent
- Visa launches CLI agent wallet: verify AI agents with a Visa card to earn OUSD — Scobleizer · 2026-11-23
- Why TLA+ users should read Lamport's original TLA paper, in the age of AI-written specs — tianyin_xu · 2026-11-23
- Ditching Mermaid: flight-board style table layouts for agent workflow UIs — thesephist · 2026-11-23
- Turn Claude Code + a free Obsidian vault into a second brain in 5 minutes — dr_cintas · 2026-11-23
- LinkedIn cuts data export in December: contacts' emails and titles dropped — thedealdirector · 2026-11-23
- Dev demos agentic mode that finds NYC events from a single natural-language prompt — thejaan · 2026-11-23