Who contributed to Local LLMs on August 8, 2026?

16 developers shipped this update, including jessegross, ParthSareen, hoyyeva, dhiltgen, ngxson, allozaur, robertomeroni, and jinzhen-lin, and 8 more.

What were the notable Local LLMs updates?

mlxrunner: add DFlash block-diffusion speculative decoding, cmd/tui: stream thinking traces in the agent TUI, and cmd/tui: restore launcher integrations menu.

Ollama ships block-diffusion speculative decoding, sglang launches kimi k3 day-0 · Local LLMs

Draft models just got faster: Ollama merged DFlash block-diffusion speculative decoding that proposes entire token sequences in one forward pass, while SGLang shipped production-ready support for Kimi K3, a 2.8T-parameter multimodal LatentMoE that routes through a latent space with 896 experts.

Ollama's DFlash implementation [1] is a fundamental architecture shift for inference optimization. Instead of predicting one token per draft call, DFlash attends over hidden states from multiple target layers and proposes a whole block in a single forward pass, borrowing the target model's embedding and output head rather than maintaining its own. The runner now handles both single-token MTP chains and multi-token block proposals with reusable rollback and adaptive depth logic. Meanwhile, Ollama's TUI evolved to stream thinking traces live [2], collapsing completed thoughts into persistent rows and letting developers reopen inline details with Ctrl+O, preserving exact token counts through live, stored, and request history. Launcher integrations now cascade cleanly [3] with overflow handling, and the Responses API finally expands namespace tool declarations [4], fixing a critical gap where namespaced function calls had no schema. SGLang v0.5.17 [5] landed day-0 Kimi K3 support with DCP, DSpark speculative decoding, chunked-prefill pipelining, and KDA-aware prefix caching stacked together. Across llama.cpp, the server gained an LRU scheduler [6] that queues model eviction and unblocks waiting requests atomically when capacity returns, the MTMD preprocessor now respects exact min/max token limits [7], and the Metal backend fixed NORM kernels for partial simdgroups [8]. vLLM hardened its Transformers multimodal path [9], patching crashes on text-only prompts to multimodal models and fixing double-tokenization of special tokens, while K3 kernel fusion [10] cut fused KV computation overhead by combining five projections into one, 4.5 to 4.6x faster. SGLang's speculative decoding now prioritizes stop tokens over length limits [11], preventing junk tokens from leaking past EOS, and HiCache's write-back policy [12] reclaims duplicated host copies first under memory pressure to maximize capacity.

Local LLMs

OLLAMA SHIPS BLOCK-DIFFUSION SPECULATIVE DECODING, SGLANG LAUNCHES KIMI K3 DAY-0

LLAMA.CPP HARDENS VULKAN STACK, VLLM CUTS ATTENTION LATENCY, SGLANG SCALES MLA

OLLAMA FIXES KV CACHE CORRUPTION, LLAMA.CPP HARDENS FILE SEARCH, VLLM KILLS MOE LEGACY CODE

OLLAMA FIXES RACE CONDITIONS AS LLAMA.CPP SHIPS BREAKING TTS CHANGES

OLLAMA REWIRES OPENAI COMPATIBILITY, LLAMA.CPP HARDENS SAMPLING

LANGCHAIN PATCHES CLAUDE FILE TOOLS AND OPENAI CONTEXT ERRORS

CACHYOS AUTOMATION PUSHES BROWSER UPDATES ACROSS THE STACK

ELIXIR PARSER GETS LEANER AS PHOENIX AUDITS DEAD CODE

MAUI LOCKS IN NET11.0, RUNTIME UNIFIES STUB MEMORY