Local LLMs — full archive
93 daily briefings · 10 reviews.
$ ls local-llm/reviews/ # the step-back reads
- week vLLM v0.32.0 lands systemone, security fixes behind it
- week vLLM 0.31.0 ships, and six security holes close
- week vLLM fixes cache_salt oracle and multimodal IPC desync
- week Five engine-fatal flaws hit vLLM, llama.cpp ships v0.5.0
- week Sparse attention lands across vLLM, SGLang, llama.cpp
- week vLLM EngineCore can be killed by crafted stop_token_ids
- month vLLM 0.30.0 lands and nineteen engine-fatal flaws close
- week Vulkan top_k radix select arrives for long rows
- week vLLM 0.28.0 lands with P2P weight sync, llama.cpp and partners broaden model reach
- month Speculative decoding consolidates across the stack as DeepSeek V4 lands
- Ollama preps Phonon voice input, vLLM drops deprecated IR knob
- Ollama fixes cloud metadata for every launcher, MLX speeds 6-16 row decode ~3x
- Ollama streams real errors, llama.cpp forwards sampling params on /v1/messages
- Five vLLM crash and forgery paths, realtime transcription lands in llama.cpp
- Ollama stops leaking canceled requests upstream, llama.cpp hardens big-endian paths
- sglang v0.5.22 lands, MLX fixes M1/M2 SDPA
- vLLM v0.32.0 lands system-one endpoint; Ollama exposes template_source
- Streamed tool calls get 200x faster in vLLM; llama.cpp learns proxies
- llama.cpp 0.7.0 ships MoE host cache, ollama reaps orphaned Windows runners
- llama.cpp lands GDN chunked kernels and SYCL IQ3 reorder
- llama.cpp server stack overflow, Ollama brings Metal kernels to CUDA
- Metal MoE fix for duplicate ids, sglang stops crashing on retracted DeepSeek-V4
- vLLM patches ZMQ socket exhaustion and CORS defaults, Ollama moves decision models into llama.cpp
- MiniCPM-V 4.7 lands in llama.cpp, vLLM fixes KV reservation under pressure
- vLLM's unauthenticated DoS surface, and a 1.9x SYCL MXFP4 lift
- vLLM drops compile-based sequence parallelism, llama.cpp fixes slot checkpoints and a silent CUDA fallback
- Ollama drops the CLI account step, MLX fixes foreign buffer ownership
- vLLM's multimodal input path is remote code execution, and MLX's sort is silently wrong
- vLLM's request-controlled caches, and Vulkan sparse attention for quantized K/V
- vLLM 0.31.0 ships DeepSeek-V4.1-Flash kernels, streaming errors stop vanishing behind DONE
- OpenVINO 2026.4.1 lands in llama.cpp, vLLM frees CUDA graph memory on sleep
- llama.cpp spec sampling goes probabilistic, vLLM trims GLM-5.3 attention overhead
- Qwen4Exp MTP lands in llama.cpp as SGLang ships v0.5.21
- MLX ports Metal gather_qmm wins to CUDA, heads-up on a qmm_sm80 corruption bug
- Ollama gates System One on declared capabilities, llama.cpp fixes a stale-UI trap
- vLLM's rejection path desyncs multimodal caches
- llama.cpp unblocks chunked reranking, vLLM patches three engine-core holes
- vLLM locks down multimodal kwargs, llama.cpp cleans up failed restores
- Tiled k-quant matmul lands in llama.cpp, SYCL gets sparse FA
- Ollama eases off typical_p, llama.cpp lands Vulkan int8 tensor cores
- llama.cpp v0.5.0 ships, vLLM discloses five engine-fatal request-handling flaws
- Ollama folds structured output into one pass, llama.cpp binds multiple addresses
- vLLM 0.30.0 lands DeepSeek-V4.1-Flash, and a min_tokens validation hole closes
- llama.cpp sanitizes invalid UTF-8 in the parser AST
- Metal MoE/SSM fusion lands, MiMo-V2.5 fp8 sharding fixed
- Ollama exposes thinking levels, llama.cpp widens Hexagon and architecture coverage
- GGUF offset corruption fixed, vLLM restores Dynamo KV metadata
- MLX engine graduates out of x/, ROCm fixes silent attention corruption
- GPU backends converge on sparse attention as vLLM wires DeepSeek V4.1 indexer
- llama.cpp 0.4.1 lands Maple 20B-A1B, Ollama drops typical_p
- SYCL memory reporting fixed, sparse-attention kernels land across vLLM and SGLang
- vLLM EngineCore can be killed by crafted stop_token_ids, llama.cpp reworks JSON schema handling
- Ollama drops its CLI agent, llama.cpp cuts build times, vLLM fixes crash under DP>1
- Ollama fixes unbounded KV memory growth, vLLM clarifies CVE timing
- vLLM 0.29.0 makes Model Runner V2 the default, llama.cpp deprecates --mmap|mlock|dio
- Ollama retries compaction after context overflow, llama.cpp fuses Vulkan kernels
- SGLang fixes TP hangs and GLM-5.2 routing
- Vulkan adds TQ1_0, Spark2.5 lands in llama.cpp
- vLLM closes scale-out multimodal handoff hole
- vLLM Fast Start maps weights zero-copy; GLMGA sampling capped
- Ollama enables speculative decoding under structured output, vLLM adds LoRA and fusion
- Error propagation hardened in ollama's MLX bindings, gemma4 gains multimodal support
- Ollama honors model generation defaults; llama.cpp and vLLM ship MoE performance fixes
- vLLM bounds validation errors, AutoRound FP8 arrives
- Vulkan top-k radix sort, HIP Q2_0 optimizations, and vLLM CUDA-graph memory fix
- llama.cpp Vulkan memory bloat fixed, vLLM deprecates PyAV
- OpenVINO backend lands Qwen3.5 on NPU, Whisper.cpp support
- LLAMA.CPP UNLEASHES QWEN3.8-FLASH-NEXT SUPPORT
- vLLM 0.28.0 LANDED WITH 584 COMMITS AND A KIMI-K3 PERFORMANCE PUSH
- OLLAMA'S MLX RUNNER NOW SPEAKS STRUCTURED OUTPUT
Never miss a Local LLMs briefing.
Get Local LLMs's week in review every Monday - and security advisories the day they land. Free, unsubscribe in one click.
One click. Every Monday: what mattered across the projects you've been reading (or our AI picks) — security news same-day. You refine anytime. Choose specific journals · Read a sample issue →