RepoJournal
Local LLMs

Local LLMs

Ollama, llama.cpp, vLLM, and SGLang - running and serving open models locally

Keep up with Local LLMs in about 3 minutes a day: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

One email a day. Unsubscribe in one click.

Pick a date

  1. Thu 23 jul
  2. Fri 24 jul
  3. Sat 25 jul
  4. Sun 26 jul
  5. Mon 27 jul
  6. Tue 28 jul
  7. Today 29 jul
Topics: Local LLMs AI / ML Full archive →

The Wire · Showcase

OLLAMA RIPS OUT IMAGE GENERATION, LLAMA.CPP SHIPS SPECULATIVE DECODING

By RepoJournal · Filed · About Local LLMs

jmorganca BBuf 王金旭 meatposes Bhavik Sharda Jeff Bolz njhill
15 people shipped this
jmorganca @jmorganca 2 cited BBuf @BBuf 2 cited 王金旭 @wjinxu 1 cited meatposes @meatposes 1 cited Bhavik Sharda @BLSharda 1 cited Jeff Bolz @jeffbolznv 1 cited njhill @njhill 1 cited BugenZhao @BugenZhao 1 cited jdebache @jdebache 1 cited mrn3088 @mrn3088 1 cited fxmarty-amd @fxmarty-amd 1 cited Liangsheng Yin @hnyls2002 1 cited YAMY1234 @YAMY1234 1 cited mickqian @mickqian 1 cited Daniel Bevenius @danbev

Ollama removed its entire MLX image generation pipeline overnight, while llama.cpp landed DSpark speculative decoding and fixed a critical SYCL memory bug affecting Arc GPUs.

Ollama [1] [2] stripped out the x/imagegen tree entirely - MLX engine, Flux2 models, C bindings, and all server routes - with plans to reintroduce image generation on a new MLX runner. This is a major cleanup that simplifies the codebase ahead of architectural changes. Meanwhile, llama.cpp merged DSpark speculative decoding [3], a new low-rank Markov head architecture layered on top of the existing DFlash drafter, with no new public APIs. The same repo fixed a critical SYCL use-after-return bug [4] in oneDNN flash-attention that corrupted the SDPA scale on Arc Pro B70 when KV cache exceeded 26k tokens. On the acceleration front, llama.cpp added chunked SSD matmul for Mamba-2 prefill [5] and restored iq4_nl quantization support to Vulkan flash-attention [6]. vLLM shipped heterogeneous KV block size support for hybrid mamba models [7], added native Kimi K3 support with Rust frontend and XTML rendering [8], optimized non-grouped topk routing to use fused kernels [9], eliminated double tensor copying in multiprocessing via zero-copy pickling [10], and fixed MXFP8 MoE dispatch on AMD MI355 [11]. SGLang refactored dead allocator code [12], updated Kimi-K3 GB200 recipes with measured 4x4 data [13], fixed Moss vision metadata validation [14], enabled multimodal prefill CUDA graphs [15], and removed stale Z-Image kernels after moving to Triton [16].

One email a day. Unsubscribe in one click.

Action items

References

  1. [1] mlx: remove experimental image generation code for now ↗ ollama/ollama
  2. [2] imagegen: remove MLX image generation code (#16615) ollama/ollama
  3. [3] spec: add DSpark speculative decoding (#25173) ggml-org/llama.cpp
  4. [4] sycl: fix use-after-return of the SDPA scale in the oneDNN flash-attention path (#25880) ggml-org/llama.cpp
  5. [5] ggml-cuda: add chunked SSD matmul for Mamba-2 prefill acceleration (#22675) ggml-org/llama.cpp
  6. [6] vulkan: add iq4_nl support back to FA (#24585) ggml-org/llama.cpp
  7. [7] [KV Connector] Support NIXL heterogeneous P/D block sizes for hybrid models ↗ vllm-project/vllm
  8. [8] [Model] Add Kimi K3 support: Rust frontend [1/2] ↗ vllm-project/vllm
  9. [9] perf: dispatch non-grouped bias-less topk routing methods to fused path ↗ vllm-project/vllm
  10. [10] [Perf] Zero-copy torch.Tensor pickling in shm_broadcast MessageQueue ↗ vllm-project/vllm
  11. [11] [MXFP8][ROCm] Fix MXFP8 MoE backend selection ↗ vllm-project/vllm
  12. [12] [Refactor] Remove dead allocator `backup_state` / `restore_state` (#32709) sgl-project/sglang
  13. [13] docs(cookbook): update Kimi-K3 GB200 recipes from measured 4x4 runs ↗ sgl-project/sglang
  14. [14] fix(vlm): reject Moss vision metadata mismatches ↗ sgl-project/sglang
  15. [15] Enable multimodal prefill BCG for VL and audio models ↗ sgl-project/sglang
  16. [16] [diffusion] refactor: remove stale kernels and dead code ↗ sgl-project/sglang

Quick answers

What shipped in Local LLMs on July 29, 2026?
Ollama removed its entire MLX image generation pipeline overnight, while llama.cpp landed DSpark speculative decoding and fixed a critical SYCL memory bug affecting Arc GPUs. In total, 115 commits, 115 pull requests, and 9 releases landed.
Who contributed to Local LLMs on July 29, 2026?
15 developers shipped this update, including jmorganca, 王金旭, Daniel Bevenius, meatposes, Bhavik Sharda, Jeff Bolz, njhill, and BugenZhao, and 7 more.
What were the notable Local LLMs updates?
mlx: remove experimental image generation code for now, imagegen: remove MLX image generation code (#16615), and spec: add DSpark speculative decoding (#25173).

More from Local LLMs

Local LLMs Local LLMs

OLLAMA REWRITES TUI PERFORMANCE LAYER AS LLAMA.CPP ARMS MEMORY TIERS

Ollama shipped three performance wins for the terminal interface while llama.cpp redefines mlock semantics and vLLM closes quant bugs across ROCm and transformer backends.

ParthSareen dhiltgen Aaron Teo Georgi Gerganov +12
llms 196 shipped 1-min read
Local LLMs Local LLMs

LLAMA.CPP SHIPS MINIMAX-M3 SPARSE ATTENTION, VLLM UNLOCKS HETEROGENEOUS PREFILL-DECODE

The inference stack gets a new model architecture and a fundamental shift in how prefill and decode can run on different hardware configurations.

github-actions[bot] dhiltgen timkhronos Pascal +10
llms 125 shipped 1-min read
Local LLMs Local LLMs

LLAMA.CPP LOCKS DOWN WASM, ADDS MCP STDIO, HEXAGON GETS IM2COL ACCELERATION

llama.cpp disabled OpenMP for Emscripten builds to fix WASM compilation, while shipping full MCP stdio support and unlocking 1.5x to 3.5x latency gains on Hexagon vision models.

reeselevine ngxson tboinovski1 angt +9
llms 139 shipped 2-min read
Local LLMs Local LLMs

OLLAMA LANDS LAGUNA SUPPORT AND CRUSHES MEMORY LEAKS WHILE SGLANG HITS V0.5.16 WITH CONFIDENCE-DRIVEN SPECULATIVE DECODING

Ollama shipped three critical performance and reliability fixes for Metal residency and concurrent access patterns, while SGL-Lang released 0.5.16 with a new speculative algorithm hitting 383.7 tok/s on DeepSeek-V4.

dhiltgen ParthSareen Xuan-Son Nguyen Johannes Gäßler +11
llms 210 shipped 2-min read

Daily updates, in your inbox

Follow Local LLMs

Keep up with Local LLMs in about 3 minutes a day: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

One email a day. Unsubscribe in one click. Read a past issue →

Elsewhere on the wire

AI Agents AI Agents

LANGCHAIN CORES SHIP, ANTHROPIC THINKING BLOCKS FIXED, OPENHANDS TIGHTENS WEBSOCKET AUTH

LangChain released two core packages with critical fixes for streaming stability and environment variable handling, while OpenHands closed a security gap that left WebSocket connections unprotected.

github-actions[bot] langchain-oss-model-profiles[bot] Functionhx ccurme +9
ai-agents 119 shipped 1-min read
CachyOS @CachyOS

KERNEL SCHEDULER UPDATES ACROSS BORE AND SCX TOOLING

CachyOS shipped BORE 6.8.0 patches across two kernel series while upgrading its sched-ext ecosystem in parallel, signaling major tuning work ahead.

Piotr Gorski Vladislav Nepogodin witchlliee Eric Naim +1
infra 83 shipped 1-min read
Elixir & Phoenix Elixir & Phoenix

ELIXIR TIGHTENS ERROR HANDLING ACROSS DATE, ECTO, AND IEXHELPERS

Date.range now surfaces ArgumentError instead of Protocol.UndefinedError when passed invalid step values, catching a class of bugs that silently fail in production code.

Alexander Gubarev Steffen Deusch lukaszsamson José Valim
elixir 14 shipped 1-min read
.NET @dotnet

MAUI CI SCANNER FIXED, MSBUILD ALIGNS TO .NET 10.0.10

A critical MAUI CI pipeline flaw that could pass builds with unscanned device tests is now patched, while MSBuild runtime dependencies ship their first 10.0.10 servicing bump.

Shane Neuville ViktorHofer dotnet-maestro sayedihashimi +4
dotnet 41 shipped 1-min read
Want every project, not just this one?