The release-to-runtime gap collapses
LFM3 reaches local runtimes immediately, evaluation shifts toward private tests and maintainer judgment, and agent infrastructure exposes more of its operating controls.
The useful unit of open AI is becoming the release plus its runtime path.
LFM3 arrived as six small text, vision, audio, omni and decision models, but the more consequential fact was how quickly the surrounding stack moved. Ollama shipped stable Apple support in the same window, while llama.cpp added model conversion, recurrent-state rollback, voices and multimodal paths. The old delay between weights and a usable local system is shrinking into the release itself. @liquidai ↗ · github-actions[bot] ↗ · github-actions[bot] ↗ · github-actions[bot] ↗
The evaluation layer is changing too. Artificial Analysis moved most of its index to private or held-out tests, and FrontierCode asks whether maintainers should merge underspecified patches. Those choices may resist contamination, but they also make benchmark governance part of the score. Vendor performance claims remain unreplicated here, Reddit collection failed provider-wide, and the Discord cache predates this window. Builders should track integration latency and evaluation assumptions alongside raw capability. @ArtificialAnlys ↗ · @cognition ↗ · @nrehiew_ ↗
- The caveat: Fast integration can still hide platform limits: Ollama’s stable LFM3 and d1 support is Apple-only for now, and several benchmark results are publisher-reported.
- What to watch: Treat runtime availability, model-specific state handling, tool semantics and evaluation governance as first-class release criteria.
X
Collection scope
Fifty-one in-window posts from bounded logged-out profile samples. Two profiles exposed no supported post payload; timelines, replies and reference-list completeness are not established.
Models
- Liquid AI releases six LFM3 variants Liquid AI’s announcement spans 1.7B text plus 2B vision, audio, omni and decision models. It claims 256K text context, eighteen-language audio and one-GPU fine-tuning; performance remains vendor-reported. @liquidai ↗
- LFM3 reaches browser inference Xenova highlights local real-time video captioning for LFM3 through Transformers.js and WebGPU. The post establishes an available browser path, not broad device performance or production reliability. @xenovacom ↗
- Ollama packages the local path Ollama says LFM3 text, vision, audio, omni and decision variants are available locally; the 4-bit text model is about 1GB. Its release notes narrow current support to Apple hardware. @ollama ↗
- Spark compresses an agent model Spark-X2.6-1.7B advertises one-million-token native context, three reasoning levels and day-zero vLLM support. Publisher figures put it near the 4B sibling on SWE-Verified and tau²-bench. @vllm_project ↗ · @SparkLLM ↗
- Ling updates its open-weight reasoning model Artificial Analysis describes Ling 3.1 Flash 1019 as a 560B-total, 25B-active MIT model with one-million-token context. Its reported score rises while output tokens per task increase more than 40%. @ArtificialAnlys ↗
Evals
- Artificial Analysis moves behind held-out tests Intelligence Index v5 says 60% of its composite now uses private datasets or held-out answers. It adds FrontierCode 2.0 and Terminal-Bench-Science while revising three existing evaluations. @ArtificialAnlys ↗
- The measured open-weight gap widens On the new index, Artificial Analysis reports released open weights led by DeepSeek V4.1 Flash 1017 and GLM-5.3 at 32, twenty-six points behind its overall leader. @ArtificialAnlys ↗
- FrontierCode asks whether code should merge FrontierCode 2.0 uses larger, underspecified tasks and asks maintainers whether patches should be merged. Its publisher reports a 45% top score, leaving substantial headroom and more subjective review. @cognition ↗ · @nrehiew_ ↗
Systems
- Snowflake treats reward hacking as a systems problem Snowflake’s research thread recommends overfitting a 160-task sandbox before scaling, isolating environments, gating raw formats and curating data; its FlashNext-RL work then scales dense recipes toward MoE training. @StasBekman ↗
- Together automates bounded GPU repair Together says cluster health checks can identify GPU failures and choose the lightest repair, with operators deciding which actions run unattended down to individual XID codes. @togethercompute ↗
- Together maps a composable open stack Together frames the open stack as model, inference, gateway or router, harness and tools. The useful claim is architectural: each layer can change without replacing the others. @togethercompute ↗
- Coding agents invert maintainer economics Ostris argues that generating a feature or fix can now take less effort than reviewing and testing a contributed patch. It is a maintainer observation, not measured ecosystem-wide evidence. @ostrisai ↗
- Epoch maps compute concentration Epoch AI estimates 72% of post-2022 AI compute went to the US and about 10% to China, using customs records. The figures come from its new chip-import explorer and methodology. @EpochAIResearch ↗
Collection scope
No current Reddit evidence. All fourteen paced Arctic Shift calls failed with HTTP 500 or 429 responses. Yesterday’s submissions and comments were not reused as new.
Discord
Collection scope
No in-window Discord messages. Approved OSNews collector caches were under fourteen hours old and retained true timestamps with cloudBrowserRead:false, but their latest captures at 14:32–14:34 UTC preceded this window’s 14:56 UTC start.
Primary sources
Collection scope
Official release feeds, publisher posts and arXiv abstracts were checked. Three HTML-only sources remain manual; feed timestamps establish discovery and publication metadata, not independent reproduction.
Local runtimes
- Ollama 0.40.12 makes LFM3 stable Ollama’s stable release supports LFM3 and d1 2B variants on Apple hardware, covering text, image, audio and omni inputs. Linux and Windows support is explicitly listed as forthcoming. github-actions[bot] ↗
- Ollama updates Apple prompt processing The same release speeds prompt processing for Gemma 4, Flint and Laguna on Apple silicon, adds DFlash2 speculative decoding on MLX, and preserves more detail in small Gemma 4 images. github-actions[bot] ↗
- llama.cpp adds LFM3 text and vision llama.cpp build 11938 adds LFM3-1.7B and LFM3-VL-2B, enables recurrent-state rollback, keeps vision depthwise convolution kernels in F32 during quantization and adds a Metal fusion baseline. github-actions[bot] ↗
- llama.cpp adds LFM3 audio and omni Build 11941 adds LFM3 Audio and Omni, keeps encoder convolutions and codecs unquantized, disables depthformer flash attention, and exposes male or female voices through CLI and server TTS. github-actions[bot] ↗
- Decision models join llama.cpp Build 11945 adds LiquidAI’s d1-2B and d1-omni-2B while keeping speech-task templates out of their GGUF files. The release entry establishes support, not decision-quality benchmarks. github-actions[bot] ↗
Agent interfaces
- llama.cpp adds a realtime endpoint Build 11946 adds conversation and transcription on /v1/realtime and fixes concurrent slots sampling another session’s logits. It also makes the realtime client use model-card prompts. github-actions[bot] ↗
- Servers expose supported reasoning levels Build 11944 probes chat templates for minimal through max reasoning levels, detects aliases and reports supported values in /props. Clients can discover capability instead of guessing template behavior. github-actions[bot] ↗
- Deferred tools enter common chat handling Build 11940 adds deferred-tool and tool-reference content parts to llama.cpp’s common chat layer. The release note records protocol support; interoperability across clients was not tested here. github-actions[bot] ↗
- Sparse cache selection is decoupled from dense masks Build 11936 constructs Qwen4exp and GLM5-next cache-cell selection independently of dense kq-mask shape, allowing dense layouts to change without rewriting sparse-attention selection. github-actions[bot] ↗
- LiteLLM trims rejected-request logs LiteLLM 1.106.5 keeps rejected-request failure logs to core request fields, tags and guardrail details, and makes the GCS logger use proxy-configured bucket credentials. The fix was backported. yuneng-berri ↗
Research
- Attention gates target prompt injection Heads of Security learns per-head gates between trusted and untrusted tokens. The paper reports RL-Hammer attack success falling from 97% on GPT-OSS-20B to 10% with its defense. Igor Shilov ↗
- Agent improvement filters noisy failures GEPA-Curator separates prompt weaknesses from harness, execution and judge noise when learning from failed rollouts. Across four environments, its authors report gains up to 20.7% over GEPA. Zack Yu ↗
- Linear attention routes with second moments MERLIN tracks mean and variance for each memory segment, then routes per-state readouts by query. Its authors report long-context gains over baselines while using less memory than some alternatives. Hee Seung Hwang ↗
- Group memory fails by preserving the wrong evidence CrossProfile finds relevance filters can re-admit topical decoys while rarest-first selection averages no better than random. Removing agent-attribution tags improves reported accuracy by 13 to 24 points. Anouar Imel ↗
- RL scheduling moves to environment-task pairs EnvSched predicts reward variance across environment-task pairs and allocates batches toward informative, underobserved interactions. On Qwen3-8B, its authors report beating random selection with one-third the training steps. Sihan Yang ↗