Local LLM Tuning, Agent Verification, and Open-Weight Releases: Today's AI Digest

Today's findings

1. #1 Get 35B-A3B MoE to 60 tok/s on 16GB cardstechnique

   Two community-tuned configs push Qwen3.6-35B-A3B to ~60 tok/s on 16GB GPUs with 256k context.

   Why it matters: MoE models are the sweet spot for local coding agents, but default settings can leave you at 3 tok/s; these settings unlock usable speed on consumer hardware.

   How to apply: Apply the three key settings: use Q4_K_M quantization, enable multi-token prediction, and adjust KV cache offload; see the linked posts for exact llama.cpp/Ollama flags.

   local llmquantizationmoeperformance

   Read more: Single 16 GB 5070 Ti running a 35B-A3B MoE at 256k context, ~55–66 tok/s — the 3 settings that took it from 3 tok/s to 60 · Qwen3.8-27B-UD-Q3_K_XL on 9070XT 16G VRAM - 96K Context

2. #2 Qwen 3.8 27B matches year-old SOTA on Aidertechnique

   Qwen 3.8 27B scores 72.9 on Aider, matching Gemini 2.5 Pro and beating Claude Opus 4 from 2025.

   Aider benchmark

   Local 27B matches year-old SOTA

   72.9

   Aider score

   Ties Gemini 2.5 Pro · beats Claude Opus 4 (2025)

   27B

   parameters

   256k

   context

   FP8

   KV cache

   A local 27B model now rivals frontier models from a year ago for coding.

   Why it matters: A local 27B model now rivals frontier models from a year ago for coding tasks, making local coding agents viable without cloud APIs.

   How to apply: Run Qwen 3.8 27B with FP8 KV cache and 256k context in vLLM; use it as your default local coding model.

   local llmcodingbenchmark

   Read more: Qwen 3.8 27B Aider score

3. #3 MiniMax H3 Fun Controlnet Union releasedrepo

   Alibaba PAI released a Controlnet Union for MiniMax H3 supporting Canny, Depth, HED, MLSD, Pose, and video inpainting.

   Alibaba PAI

   Controlnet Union for MiniMax H3

   6.8GB model on HuggingFace — precise structural control in ComfyUI.

   Why it matters: Gives open-source video generation precise structural control, enabling consistent character and motion in ComfyUI workflows.

   How to apply: Download the 6.8GB model from HuggingFace and wire it into your H3 workflow; community examples are in the round-up.

   videocontrolnetopen-sourcecomfyui

   Read more: MiniMax-H3 Fun Controlnet Union released · A quick Minimax H3 news round-up - 24th August 2026

4. #4 Disable pinned memory to stop ComfyUI RAM crashestip

   Launching ComfyUI with --disable-pinned-memory eliminated hostbuffer errors and cut RAM usage dramatically on a 5090.

   ComfyUI memory fix

   One flag stops RAM crashes

   Pinned memory ON

   • Hostbuffer errors
   • ~25GB RAM

   → 25GB RAM saved

   Pinned memory OFF

   • Stable runs
   • Low RAM

   Add --disable-pinned-memory to your launch command.

   Why it matters: Pinned memory can eat 25GB of system RAM and cause random crashes during local generation; this flag is a one-line fix.

   How to apply: Add --disable-pinned-memory to your ComfyUI launch command; test on your own hardware to confirm stability.

   comfyuimemorylocal

   Read more: ComfyUI was eating my RAM and causing crashes, this fixed it

5. #5 OpenPitStop: open-source referee for AI coding agentstool

   OpenPitStop is an independent checker that verifies whether a coding agent's fixes actually hold.

   Open-source tool

   OpenPitStop: the referee for AI coding agents

   1

   Agent fix

   modifies code

   2

   OpenPitStop

   verifies changes

   Independent check

   3

   Merge

   if clean

   Run alongside Claude Code to catch regressions before merge.

   Why it matters: Coding agents can modify large codebases, but nobody checks the work; this tool adds a verification layer to catch regressions.

   How to apply: Run OpenPitStop alongside your agent (e.g., Claude Code) to automatically test and validate changes before merge.

   agentscodingverificationopen-source

   Read more: How do you know when an AI coding agent is actually done?

6. #6 Disable Claude Code's model downgrade to avoid destructive editstip

   Claude Code can silently downgrade to Haiku during high load, leading to sloppy edits; disable it in settings.

   Why it matters: A downgraded model can delete files or make bad changes without warning, costing time and data.

   How to apply: In Claude Code settings, turn off 'downgrade model' to keep Opus/Sonnet quality; also consider registering a free fallback model.

   claudeclaude codesafety

   Read more: Pro tip DISABLE DOWNGRADE MODEL

7. #7 Preserve rejected decisions across Claude Code compactiontip

   Compaction keeps what you built but forgets what you rejected; write a handoff note before compacting to avoid re-exploring dead ends.

   Claude Code tip

   Preserve rejected decisions across compaction

   Without handoff note

   • Rejected decisions forgotten
   • Agent re-explores dead ends
   • Wasted time

   vs

   With handoff note

   • Rejected decisions preserved
   • Agent avoids dead ends
   • Efficient continuation

   Write a handoff note before /compact

   Add a 'decisions made and rejected' section to context or CLAUDE.md.

   Why it matters: Long sessions lose the reasoning behind discarded approaches, causing the agent to repeat failed ideas.

   How to apply: Before /compact, add a short 'decisions made and rejected' section to your context or CLAUDE.md.

   claudecontextmemory

   Read more: Compaction keeps what we built and forgets what we rejected

8. #8 Scaling parallel Claude Code agents: the bottleneck chaintip

   Running 30 parallel agents revealed a chain of bottlenecks: cognition → billing → CPU → OS → physics; each fix reveals the next.

   Why it matters: If you parallelize agents, expect limits beyond just model quality; plan for infrastructure and attention constraints.

   How to apply: Start with 5-6 agents, monitor CPU/thermal, and use tools like tmux or dedicated UIs to manage context; upgrade hardware as needed.

   agentsclaudescaling

   Read more: I kept adding parallel agents until my Mac shut down from heat. The bottleneck moved down a layer every time I fixed it.

9. #9 Track success/failure counts for learned agent workflowstechnique

   Store a success_count and fail_count on every learned workflow so agents don't blindly follow a version that never worked.

   Why it matters: Agent memory often keeps the newest workflow, not the most reliable; tracking outcomes prevents regression.

   How to apply: Add version, success_count, fail_count to your workflow definitions and have the agent consult them before execution.

   agentsmemoryreliability

   Read more: Does anything in your setup track whether a learned workflow actually worked?

10. #10 AQuA v2: agent memory of rejected experiments and bounded model searchpaper

    AQuA v2 shows agents that remember rejected experiments and keep the evaluator fixed produce more reliable research loops.

    Why it matters: This paper offers a concrete pattern for agent memory and self-improvement that you can apply to your own agent pipelines.

    How to apply: Implement a persistent state that records both accepted and rejected evidence, and restrict agent changes to bounded configuration spaces.

    agentsresearchmemorypaper

    Read more: A research agent's most useful memory may be the experiment it rejected · This model-search agent can change architecture, loss, sampler or optimizer—but not the evaluator

11. #11 Ling-3.0 opens six base checkpoints across three training stagesrepo

    Ant Group released six Ling-3.0 base checkpoints (tiny and flash, each with pretrained, mid-trained, and WSM-merged versions).

    Ant Group · Ling-3.0

    Six base checkpoints across two sizes and three training stages

    tiny · pretrained base

    tiny · mid-trained base

    tiny · WSM-merged base

    flash · pretrained base

    flash · mid-trained base

    flash · WSM-merged base

    Pretrained Mid-trained WSM-merged

    Stage labels matter for reproducibility and downstream fine-tuning.

    Why it matters: Open-weight base models give you fine-tuning freedom; the stage labels matter for reproducibility and downstream training.

    How to apply: Pick the right stage for your use case (e.g., mid-trained for chat fine-tunes) and record exact repo/revision in your registry.

    open-weightsfine-tuningllm

    Read more: Ling-3.0 just opened six base checkpoints across three training stages · Ling-3.0 has six base checkpoints. A family name is not an artifact ID.

12. #12 bf16-to-f16 GGUF conversion for Strix Halotool

    A community script converts bf16 GGUF models to f16 for Strix Halo hardware, enabling local runs on AMD APUs.

    Why it matters: Many models ship bf16, which some AMD iGPUs can't handle; this conversion makes them runnable locally.

    How to apply: Use the script from GitHub to convert your GGUF files before loading in llama.cpp or Ollama on Strix Halo.

    ggufquantizationlocalamd

    Read more: instead of begging ai labs cooks I was invited to make my own gguf bf16 to f16 conversion setup for strix halo models fitting and despite a couple of showstoppers along the way it was a smooth ride. Here is the result, enjoy: https://github.com/littlett77/gguf-bf16-to-f16-conversion-script-for-strix