8BitsForge Lab NotesExact builds, reproducible recipes, measurements, caveats, and honest verdicts from local model, agent, and systems experiments.https://8bitsforge.com/enSqueezing Qwen3.8 Until It Cried: A 46.9% Speedup and One Optimization Too Farhttps://8bitsforge.com/techblog/how-a-baked-mtp-head-made-qwen3-8-46-9-percent-faster/https://8bitsforge.com/techblog/how-a-baked-mtp-head-made-qwen3-8-46-9-percent-faster/A 46.9% Qwen3.8 speedup, a production qualification, and the tempting 64K shortcut that failed when the benchmark got serious.Mon, 14 Sep 2026 14:31:21 GMTLocal AIQwen3.8vLLM XPUArc Pro B70MTPGPTQBenchmarkingModel RouterUpdating OpenClaw Is Always a Rollercoaster Ride, Never Boringhttps://8bitsforge.com/techblog/updating-openclaw-is-always-a-rollercoaster-ride-never-boring/https://8bitsforge.com/techblog/updating-openclaw-is-always-a-rollercoaster-ride-never-boring/A healthy container, an unavailable Codex harness, dependencies in the wrong neighbourhood, and the rollback lever that became our favourite feature.Mon, 07 Sep 2026 17:06:09 GMTOpenClawOpenClawDockerCodexRollbackReliabilityAn A/B OpenClaw Upgrade with an Actual Escape Routehttps://8bitsforge.com/techblog/an-ab-openclaw-upgrade-with-an-actual-escape-route/https://8bitsforge.com/techblog/an-ab-openclaw-upgrade-with-an-actual-escape-route/How we promoted OpenClaw 2026.8.2 from a quiesced clone, kept 2026.8.1 ready for rollback, and removed several safety checks that were only pretending to help.Thu, 03 Sep 2026 18:15:26 GMTOpenClawOpenClawContainersA/B DeploymentsRollbackReliability210K Context on a 32 GB Arc Pro B70https://8bitsforge.com/techblog/210k-context-on-a-32gb-arc-pro-b70/https://8bitsforge.com/techblog/210k-context-on-a-32gb-arc-pro-b70/Three identical cold starts, 179,525 prompt tokens, and a rollback regex that briefly became the main character.Mon, 31 Aug 2026 00:00:00 GMTLocal AIQwen3.8vLLM XPUArc Pro B70Long ContextModel RouterHow ubatch=1024 Won the Benchmark and Lost the Ability to Speakhttps://8bitsforge.com/techblog/how-ubatch-1024-won-the-benchmark-and-lost-the-ability-to-speak/https://8bitsforge.com/techblog/how-ubatch-1024-won-the-benchmark-and-lost-the-ability-to-speak/A 6% prefill improvement, five failed conversations, and one tiny silicon cult chanting the unused32 token.Mon, 31 Aug 2026 00:00:00 GMTLocal AIGemma 4llama.cppVulkanMTPBenchmarkingQwen3.8 Flash-Next on One Arc Pro B70https://8bitsforge.com/techblog/qwen3-8-flash-next-on-one-arc-pro-b70/https://8bitsforge.com/techblog/qwen3-8-flash-next-on-one-arc-pro-b70/Fifteen tokens per second, one extremely narrow memory edge, and several acceleration ideas asked to leave the laboratory.Mon, 31 Aug 2026 00:00:00 GMTLocal AIQwen3.8Flash-Nextllama.cppVulkanArc Pro B70Benchmarking