Who contributed to Local LLMs on August 12, 2026?

15 developers shipped this update, including dhiltgen, Jesse Gross, hoyyeva, Xuan-Son Nguyen, github-actions[bot], stefankoncarevic, mgoin, and Luosuu, and 7 more.

What were the notable Local LLMs updates?

api: stop applying repeat_penalty 1.1 to models that don't set one, nemotron_h: support the Nemotron 3.5 prompt layout, and mtmd: support pocket-tts (#26871).

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 CUTS REPEAT PENALTY BLOAT, LLAMA.CPP SHIPS POCKET-TTS SUPPORT

By RepoJournal · Filed 07:02 UTC on August 12, 2026 · About Local LLMs

15 people shipped this

dhiltgen @dhiltgen 1 cited

Jesse Gross @jessegross 1 cited

Xuan-Son Nguyen @ngxson 1 cited

stefankoncarevic @stefankoncarevic 1 cited

mgoin @mgoin 1 cited

askliar @askliar 1 cited

mickqian @mickqian 1 cited

jybsuper @jybsuper 1 cited

hoyyeva @hoyyeva

github-actions[bot] @github-actions[bot]

Liangsheng Yin @hnyls2002

Ollama kills a decade of default penalty cruft that was drowning every model, while llama.cpp gains multimodal audio chops and vLLM fixes ROCm speculative decoding corruption.

Ollama dropped the 1.1 repeat penalty default that was silently tanking generation quality across its entire library [1]. That 1.1 figure was never endorsed by any model maker - llama.cpp ditched it in 2024, and vLLM, SGLang, and transformers apply nothing - so Ollama was the outlier imposing a penalty most creators explicitly rejected. On the llama.cpp side, a fresh wave of releases shipped multimodal audio support through pocket-tts integration [3], a system that replaces depthwise upsampling convolutions with GEMM operations to slash kernel-launch overhead on the decoder. Ollama also landed Nemotron 3.5 parser and renderer registration with exact Jinja parity enforcement [2], ensuring models created with the 3.5 prompt layout cannot silently regress to the 3 renderer. Over on vLLM, three critical fixes hit: hybrid models on MI300 now allocate separate pages for KV-first attention blocks to stop DSpark speculative decoding from producing garbage [4], a generalized KV block zeroing mechanism now catches all attention types including sliding-window and chunked-local [5], and FlashInfer XQA decode landed on SM12x chips through dedicated API paths [6]. SGLang tightened DiT layerwise offload policy away from fragile class-name checks toward explicit per-model modes, restoring low-memory MOVA offload [7] and refocused LoRA regression coverage [8]. Across all four repos: 125 commits, 125 PRs, and 8 releases shipped in this window.

Action items

→ Upgrade Ollama immediately to pick up the repeat penalty fix - this affects output quality on every model ollama/ollama [immediate]
→ If running vLLM on MI300 with speculative decoding, apply the KV page isolation patch before next inference run vllm-project/vllm [immediate]
→ Review llama.cpp pocket-tts release notes if you run multimodal audio workloads ggml-org/llama.cpp [plan]
→ Monitor SGLang's DiT auto-offload config if you deploy large vision models on MOVA sgl-project/sglang [monitor]

References

[1] api: stop applying repeat_penalty 1.1 to models that don't set one ollama/ollama
[2] nemotron_h: support the Nemotron 3.5 prompt layout ↗ ollama/ollama
[3] mtmd: support pocket-tts (#26871) ggml-org/llama.cpp
[4] [Bugfix][ROCm] Give KV-first attention blocks their own page in hybrid models ↗ vllm-project/vllm
[5] [Bugfix] Generalize KV block zeroing to `AttentionSpec` ↗ vllm-project/vllm
[6] [Attention] Add FlashInfer XQA decode support on SM12x ↗ vllm-project/vllm
[7] Fix model-driven DiT layerwise offload auto policy ↗ sgl-project/sglang
[8] Refocus LoRA tests on regression coverage ↗ sgl-project/sglang

Quick answers

What shipped in Local LLMs on August 12, 2026?: Ollama kills a decade of default penalty cruft that was drowning every model, while llama.cpp gains multimodal audio chops and vLLM fixes ROCm speculative decoding corruption. In total, 125 commits, 125 pull requests, and 8 releases landed.
Who contributed to Local LLMs on August 12, 2026?: 15 developers shipped this update, including dhiltgen, Jesse Gross, hoyyeva, Xuan-Son Nguyen, github-actions[bot], stefankoncarevic, mgoin, and Luosuu, and 7 more.
What were the notable Local LLMs updates?: api: stop applying repeat_penalty 1.1 to models that don't set one, nemotron_h: support the Nemotron 3.5 prompt layout, and mtmd: support pocket-tts (#26871).

OLLAMA PATCHES CRITICAL SSRF VULNERABILITY IN MANIFEST HANDLING

A rogue OCI registry could exploit duplicate digest collisions in Ollama's blob verification to exfiltrate data from internal endpoints.

+13

llms 249 shipped

Local LLMs Aug 11

OLLAMA PATCHES CRITICAL PARSER BUG, NEMOTRON 3 LANDS WITH SPECULATIVE DECODING

Ollama shipped v0.32.7 with a critical fix for boundary token corruption in model invocation, while full MLX support for Nemotron 3 Nano Omni brings self-draft speculation to inference without a separate draft model.

+11

llms 271 shipped

Local LLMs Aug 10

OLLAMA SHIPS VISION SUPPORT ACROSS MLX AND QWEN3.5 AS LLAMA.CPP FIXES DISPATCH BUGS

Ollama's MLX runner now processes images natively, preserving prefix caching and speculative decoding while rejecting unsupported requests with clear errors.

llms 157 shipped

Local LLMs Aug 9

OLLAMA CUTS MODEL CONFIG BLOAT 100KB+ PER MODEL, LLAMA.CPP SHIPS TOOL ISOLATION AND CUDA FUSION GAINS

Ollama eliminated redundant manifest data ballooning config blobs on safetensors models, while llama.cpp landed docker-backed tool isolation and fused three GPU kernels into one.

llms 102 shipped

Elsewhere on the wire

AI Agents Aug 13

ANTHROPIC INTEGRATION FIXES TOOL BLOCKS, OPENHANDS UNBLOCKS CI, CLINE PATCHES COMPOSER WEDGE

LangChain's Anthropic integration corrected tool result handling and model profiles [ref:1], while OpenHands fixed a critical test infrastructure failure blocking all PRs and shipped 1.13.0 with conversation archives [ref:6][ref:9].