Who contributed to Local LLMs on August 5, 2026?

15 developers shipped this update, including dhiltgen, github-actions[bot], Xuan-Son Nguyen, Evan Huus, Caleb DeLeeuw, Pranav Uttarkar, zhou9402, and connorcarpenter15, and 7 more.

What were the notable Local LLMs updates?

progress: fix data races on ticker, states, spinner, and bar state, v0.32.6, and mtmd: support Qwen3-TTS (note: breaking change to llama-tts binary) (#26254).

Local LLMs

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

Pick a date

Topics: Local LLMs AI / ML Full archive →

The Wire · Showcase

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

By RepoJournal · Filed 07:02 UTC on August 5, 2026 · About Local LLMs

15 people shipped this

dhiltgen @dhiltgen 1 cited

github-actions[bot] @github-actions[bot] 1 cited

Xuan-Son Nguyen @ngxson 1 cited

Caleb DeLeeuw @SolshineCode 1 cited

zhou9402 @zhou9402 1 cited

connorcarpenter15 @connorcarpenter15 1 cited

njhill @njhill 1 cited

merrymercy @merrymercy 1 cited

hnyls2002 @hnyls2002 1 cited

BBuf @BBuf 1 cited

Evan Huus @eapache

Pranav Uttarkar @PranavUttarkar

lk-chen @lk-chen

mgoin @mgoin

Sugar920 @Sugar920

Ollama patched critical data races across its progress tracking system while llama.cpp introduces incompatible Qwen3-TTS support, and vLLM lands kernel optimizations that cut latency on specialized hardware.

Ollama's progress module had multiple goroutine synchronization bugs where NewProgress.start() wrote p.ticker while stop() read and cleared it with no locking, and Spinner/Bar state was read outside locks during concurrent operations [1]. The fix lands as v0.32.6 brings OpenAI-compatible streaming format to /v1/chat/completions, matching chunk structure with role-only on first chunk and finish_reason on its own message, plus usage in a separate chunk [2]. Llama.cpp shipped a breaking change to its llama-tts binary with Qwen3-TTS multimodal support [3], and separately fixed a tokenizer import regression where bytes_to_unicode was dropped from the transformers library but had already been copied into convert_slow_tokenizer [4]. VLLm added SM100 CuTeDSL kernel for fused query preprocessing in DSA sparse attention [5], extended its Rust gRPC frontend to handle multimodal image inputs via HTTP URLs and data URIs [6], and explicitly clamped PyTorch's CPU thread pool to single-threaded serving to prevent OpenMP spin-waiting from stealing cycles from the engine loop [7]. SGLang reduced startup log noise by consolidating eight redundant MoE warnings into one targeted log line [8], fixed VLM test image placeholders to use the model's own chat template instead of a frontend-only registry [9], and shipped FLUX.2 VAE decoder fast paths reducing inference time from 437.2 ms to 168.5 ms on H200 for 2048^2 images [10].

Action items

→ If running Ollama in production with streaming chat completions, upgrade to v0.32.6 for correctness and OpenAI format compatibility ollama/ollama [plan]
→ Review if you have local llama.cpp TTS integrations before pulling the latest llama-tts binary changes ggml-org/llama.cpp [monitor]
→ For vLLM users on SM100 or H200 hardware running diffusion workloads, profile with the new kernel optimizations vllm-project/vllm [monitor]
→ If running SGLang with VLM tests, pull the chat template fix to avoid silent image drops sgl-project/sglang [plan]

References

[1] progress: fix data races on ticker, states, spinner, and bar state ↗ ollama/ollama
[2] v0.32.6 ↗ ollama/ollama
[3] mtmd: support Qwen3-TTS (note: breaking change to llama-tts binary) (#26254) ggml-org/llama.cpp
[4] convert : import bytes_to_unicode from convert_slow_tokenizer (#26217) ggml-org/llama.cpp
[5] [Kernel][SM100] Add a CuTeDSL fused query kernel ↗ vllm-project/vllm
[6] [Rust Frontend][gRPC] Add multimodal image inference ↗ vllm-project/vllm
[7] [Core] Explicitly manage torch CPU threads in workers ↗ vllm-project/vllm
[8] Reduce startup log noise and fix Dynamo / CUDA-graph edge cases ↗ sgl-project/sglang
[9] [Fix] Resolve VLM test image placeholders from the model's own chat template ↗ sgl-project/sglang
[10] [diffusion] FLUX.2 VAE decoder fast path behind quality=high (H200: 1024^2 97.6->29.2 ms, 2048^2 437.2->168.5 ms) ↗ sgl-project/sglang

Quick answers

What shipped in Local LLMs on August 5, 2026?: Ollama patched critical data races across its progress tracking system while llama.cpp introduces incompatible Qwen3-TTS support, and vLLM lands kernel optimizations that cut latency on specialized hardware. In total, 118 commits, 117 pull requests, and 11 releases landed.
Who contributed to Local LLMs on August 5, 2026?: 15 developers shipped this update, including dhiltgen, github-actions[bot], Xuan-Son Nguyen, Evan Huus, Caleb DeLeeuw, Pranav Uttarkar, zhou9402, and connorcarpenter15, and 7 more.
What were the notable Local LLMs updates?: progress: fix data races on ticker, states, spinner, and bar state, v0.32.6, and mtmd: support Qwen3-TTS (note: breaking change to llama-tts binary) (#26254).

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

Ollama patched a critical cache corruption bug that corrupts lazy snapshots during multi-request prefills, while llama.cpp hardened Windows junction traversal and vLLM removed 14 hardcoded MoE shards across quantization methods.

+10

llms 240 shipped

Local LLMs Aug 4

OLLAMA REWIRES OPENAI COMPATIBILITY, LLAMA.CPP HARDENS SAMPLING

Ollama reworked its entire streaming protocol to match OpenAI's wire format chunk-for-chunk, unblocking drop-in replacement deployments.

+11

llms 230 shipped

Local LLMs Aug 3

DEEPSEEK V4 INFERENCE GETS 50% SPEEDUP; LLAMA.CPP AND VLLM BOTH SHIPPING OPTIMIZED PATHS

DeepSeek V4 speculative decode just went from nice-to-have to standard: llama.cpp landed MTP and DSpark support with roughly 50% speedup [ref:3], vLLM hardened its KV cache coordinator for the same models [ref:7], and Metal got native hyper-connection ops [ref:5].

llms 148 shipped

Local LLMs Aug 2

LLAMA.CPP SHIPS FOUR BUILDS IN 24 HOURS AS VLLM AND SGLANG RACE ON SEQUENCE PARALLELISM

The local LLM stack is moving fast: llama.cpp pushed builds b10219-b10223 with reasoning persistence and security updates, while vLLM landed sequence parallelism for DeepSeek V4 and sglang unified its radix cache for multi-turn agentic work.

llms 82 shipped

Elsewhere on the wire

AI Agents Aug 6

LANGCHAIN HARDENS ANTHROPIC TOOL CALLING AND GIT CREDENTIALS AS OPENHANDS SHIPS 1.10.0

LangChain fixed a critical credential leak in evaluation runs while Anthropic tool schemas get proper fallback handling, as three major agent frameworks push releases across the ecosystem.

ai-agents 54 shipped

@CachyOS Aug 6

WIKI TRANSLATION ROLLBACK CHAOS, MIRROR STACK HOLDING STEADY

CachyOS wiki maintainers are caught in a revert loop over German translations while the status infrastructure logs 55 commits without incident.

infra 57 shipped

Elixir & Phoenix Aug 6

ELIXIR BUMPS ZIZMOR SECURITY SCANNER TO 1.28.0

The Elixir repo updated its CI security action overnight, pulling in the latest zizmor static analysis tooling.

elixir 1 shipped

@dotnet Aug 6

DOTNET FOUNDATION PULLS iOS METRICS INTO FOCUS AS RUNTIME HARDENS HTTP PARSING

Performance team fixed iOS size reporting to measure only distributable packages, while runtime clamped down on RFC 9112 compliance for chunked encoding.

dotnet 72 shipped

Want every project, not just this one?

Follow Local LLMs