llama.cpp adds Q1_0, Q2_0 quant support for Hexagon NPUs
Snapdragon devices get another notch of compression for on-device inference
From llama.cpp releases/Written from the primary source/
The publication AI reads
Every model release, sourced field by field.
Exact parameter counts, architectural context length, licence and agent capability — each traced to the endpoint it came from, and readable by your agent over MCP.
422releases on record
328exact parameter counts
175chat templates read
119show tool calling
Leader · our view
Read the argumentBenchmarks measure harnesses, not the systems you'll run
Four unrelated evals this week all erred the same direction: optimistic
The register
Open the register422 releases, every field sourced
● evidence in chat template○ template read, none found◇ no template published
| Model | Parameters | Context | Tools | Think | Vision | Licence | Released |
|---|---|---|---|---|---|---|---|
| deepseek-aiDeepSeek-V4.1-Flash-1017 | 763.2B | 1M | ◇ | ◇ | ◇ | mit | 2026-10-17 |
| nvidiaNVIDIA-Nemotron-Labs-Dory1.0-4B-BF16 | 4.0B | 256K | ● | ● | ○ | other | 2026-10-16 |
| Comfy-OrgMOSS-Image | — | — | ◇ | ◇ | ◇ | apache-2.0 | 2026-10-16 |
| tencentHy-Memory-30B-A3B | 30.1B | 256K | ● | ● | ○ | apache-2.0 | 2026-10-15 |
| black-forest-labsflux-3-image-klein-comfy | — | — | ◇ | ◇ | ◇ | other | 2026-10-14 |
| mistralaiTimestral-209M | 209M | — | ◇ | ◇ | ◇ | other | 2026-10-14 |
Analysis · monthly
Read the analysisWhen open models learned to act
Tool calling took three years to become universal. Reasoning took two. Measured from every model’s own chat template.
Point your agent at it
A remote MCP server. Nothing to install, no key.
https://theneuralnet.com/api/mcpModels··Transformers releases
Transformers v5.20.1 adds Reflection AI's Beam model
A coding-focused model from Reflection AI lands amid a run of patch fixes
Safety··arXiv cs.CL
LLMs infer ethnicity from names more accurately than BISG
A withdrawn paper shows models can profile race without training data
Infrastructure··llama.cpp releases
llama.cpp b11690 trims memory copies in speculative decoding cache
A single-line fix, but it is the kind that adds up at scale
Infrastructure··arXiv cs.LG
Meta details how it made AMD GPUs production-ready for training
Porting recommendation models from Nvidia to AMD took more than a driver swap
Models··Hugging Face
IBM packs chat, decisions and speech into one Granite endpoint
Granite Switch 4.1 adds adapters, not GPUs, to cover three task types
Models··Hugging Face
Swiss AI releases Apertus 1.5 with vision, audio and longer context
The open-weight 8B and 70B models add multimodality, but still trail frontier rivals
Infrastructure··llama.cpp releases
llama.cpp v0.7.0 adds MoE GPU cache, new default server port
A new expert cache and port change reshape how MoE models get served
Models··Hugging Face
Meta releases Muse Spark Open, an 816B-parameter MoE model
Open weights, four modalities, a million-token window — but 1.5TB to serve
Models··Transformers releases
Transformers v5.20.0 adds Meta's 816B-parameter Muse Spark Open
Seven new model families land alongside breaking changes to cache and config APIs
Models··NVIDIA
Nvidia expands Nemotron family with Super VL and two Bolt models
Open weights now cover servers down to 4GB edge devices
Infrastructure··llama.cpp releases
llama.cpp adds BF16 flash attention kernel for Metal backend
A small patch note for a release that still runs on every Apple GPU
Safety··arXiv cs.LG
Study: signed LLM weights don't verify the quantized files you run
Across 198 popular GGUF repos, zero could be independently re-derived
Safety··arXiv cs.CL
Study: LLM safety refusals flip 18-28% across seeds, temperature
One prompt, many answers — single-shot safety tests may be measuring noise
Models··Hugging Face Hub
Liquid AI ships LFM2.5-24B-A2B, its biggest on-device agent model
A 24B mixture-of-experts model that uses just 2B active parameters per token
Research··Hugging Face
Ai2 updates climate emulator to cut extreme-weather simulation cost 300x
HiRO-ACEv2 matches a storm-resolving model's extremes but can't yet handle a warmer world
Infrastructure··AWS Machine Learning
AWS details MoE kernel fixes for SageMaker, upstreams them to vLLM
Three upstreamed vLLM fixes squeeze more throughput from Hopper GPUs
Models··Google DeepMind
Google DeepMind launches Gemini 4 Argon for robotics
A physical-reasoning model that writes its own simulations before touching hardware
Infrastructure··llama.cpp releases
llama.cpp adds image-embedding support for Qwen3.8 Flash Next
A single pull request, the release notes offer little else
Models··Google AI
Google launches Gemini 4 Argon, its new frontier model
Top spot on the Vals Index comes with a per-token price tag to match
Safety··arXiv cs.AI
Researchers trace where LLMs decide to refuse requests
Attention spots danger; MLP layers decide whether to say no
Infrastructure··arXiv cs.CL
Study finds weight order matters more than value in LLM compression
Researchers compress models to 16–64 shared values and find rank is king
Models··Hugging Face Hub
OpenBMB ships quantized MiniCPM-V 4.7, its first MoE vision model
A 35-billion-parameter model that runs at 13+ tokens a second on an iPhone
Research··arXiv cs.AI
Study: API benchmarks don't match chatbot app performance
Researchers say the gap can equal a full model version downgrade
Infrastructure··llama.cpp releases
llama.cpp adds runtime support for Prism Bonsai 2 27B
A pre-release build lets the open-source engine load the Hadamard-based model
Infrastructure··llama.cpp releases
llama.cpp patches server bug that corrupted busy slot's context
A slot-pinning bug could silently swap context mid-generation
Safety··arXiv cs.AI
Paper: AI agent security tests breached real systems in 2026
A study says OpenAI, Anthropic and Google evaluations escaped test scope. It proposes a fix.
Products··OpenAI
Asana cuts browser-agent model costs 76x using GPT-6.1 Sol
GPT-6 Astra redesigned caching in Codex. The per-run cost of Asana's StackAI agent fell from $36 to 47 cents.
Infrastructure··arXiv cs.AI
Paper: KV-cache tool gives LLMs a 50M-token memory on disk
galahad-kv reloads stored key-value state instead of recomputing it, cutting cost on one H100.
Models··Hugging Face Hub
Quantized GLM-5.3-Flash lands on Hugging Face for DGX Spark
A community NVFP4 quantization of Zai's GLM-5.3-Flash is tuned to run on Nvidia's GB10-based DGX Spark desktop box.
Safety··arXiv cs.CL
Study finds LLaMA-2 safety hinges on 0.19% of its weights
Researchers say a tiny, localised edit to one MLP layer strips safety guardrails while leaving accuracy almost untouched.
Products··AWS Machine Learning
AWS shows agents paying per API call via Bedrock AgentCore payments
Incarna cut x402 payment integration from months to days, AWS says, using a managed wallet and spending-limit service with inference router BlockRun.
Safety··OpenAI
OpenAI bans Russian, Iranian AI influence ops using fake media fronts
One Russian operation reached the highest severity tier OpenAI has recorded since it began reporting disruptions.
Infrastructure··llama.cpp releases
llama.cpp tunes CUDA kernel for Gated DeltaNet state updates
Release b11494 defaults GDN state updates to four columns per warp, a CUDA scheduling change in PR #30087.
Infrastructure··llama.cpp releases
llama.cpp adds support for LiquidAI's d1-3B decision model
The b11483 release renamed internally to lfm2-d1 brings a specialised "decision" model type to the inference engine.
Infrastructure··arXiv cs.AI
Quantization's effect on agent tool-failure recovery flips with eval design
A single-author study finds 8-bit vs 4-bit comparisons reverse depending on prompt, task set and parsing strictness.
Infrastructure··llama.cpp releases
llama.cpp extends Metal fast-path matmul to more quant types
The b11476 release widens a row-batched kernel to BF16, Q2_K, MXFP4 and IQ types.
Products··AWS Machine Learning
Claude Haiku 5.5 lands on Amazon Bedrock
A 75% price cut and new effort controls, built to run as a subagent under Opus 5.5.
Models··Transformers releases
Transformers v5.19.0 adds EmbeddingGemma 2, ships breaking MoE changes
Hugging Face's release adds Google's multimodal embedding model and changes how MoE routers and continuous batching behave.
Infrastructure··llama.cpp releases
llama.cpp adds BF16 support for XIELU on CUDA
Release b11457 closes a gap in the CUDA kernel, letting XIELU activations run in bfloat16 on Nvidia GPUs.
Research··OpenAI
OpenAI publishes new math results from an internal frontier model
Lean proofs and ChatGPT Pro compute estimates follow advice from an outside math-AI group.
Models··Google DeepMind
Google releases EmbeddingGemma 2, a 740M multimodal embedding model
Apache 2.0 model unifies text, image, audio and video for on-device search and RAG
Infrastructure··llama.cpp releases
llama.cpp adds support for K2 Horizon dense and MoVA models
Release b11454 brings a new converter, tokenizer and chat template fixes for the K2 Horizon model family.
Infrastructure··llama.cpp releases
llama.cpp patches data race in shared k-pool sequences
Release b11435 fixes a CPU-backend race affecting hybrid models with shared sequence caches.
Policy··OpenAI
OpenAI rolls out text watermarking to meet EU AI Act rules
It defaults on for ChatGPT and Codex, but only in the EU.
Infrastructure··llama.cpp releases
llama.cpp v0.6.0 adds new batch API and five decision models
A server endpoint now runs dedicated pipelines for decision-style models.
Research··arXiv cs.AI
Benchmark finds LLM-generated GPU kernels fail in production
FastKernels shows kernel-level speedups mostly vanish once agents' code runs inside real models.
Models··Hugging Face Hub
Aleph Alpha releases Kolibri-1, a German-English MoE reasoning model
78B parameters, 3.46B active. Apache 2.0. A sovereignty bet dressed as efficiency.
Infrastructure··arXiv cs.LG
New method quantizes Llama 3.2 11B Vision to 2.7 bits for phones
Researchers compress an 11B-parameter vision-language model to 3.7GB for Arm CPUs without needing the original training setup.
Infrastructure··llama.cpp releases
llama.cpp's WebGPU backend adds quantised matrix-vector support
Release b11406 extends WebGPU's MMVQ path to six more quantisation formats, including MXFP4.
Infrastructure··vLLM releases
vLLM 0.31.0 adds fast-restart weight caching and new DeepSeek optimisations
The release adds a preload daemon that keeps quantized weights resident across restarts, plus deep DeepSeek-V4.1-Flash tuning.
Models··Hugging Face
Ai2 open-sources AstaBrief, an 8B model for scientific report writing
It writes cited reports 3.5x faster than Ai2's Claude pipeline. Anyone can run it locally.
Infrastructure··arXiv cs.LG
New method cuts memory for hybrid-attention model states by 69%
DAMP quantises recurrent states in Gated DeltaNet and Kimi Delta Attention. Plain INT4 wrecks their accuracy; this doesn't.
Safety··arXiv cs.LG
Study: harmful-meme failures in VLMs are a routing problem
Researchers find Gemma-3 and Qwen3.5 often represent harmful-meme evidence internally but fail to route it to their outputs.
Research··arXiv cs.LG
Study: post-training expands, not just compresses, hard reasoning
A diagnostic probe finds post-training raises the solvable ceiling on hard math problems, not just the odds of sampling a known answer.
Research··arXiv cs.AI
Study: French costs up to 58% more tokens than English across 7 LLMs
A paper measures a persistent "language tax" in 2026 tokenizers, and tests a French-optimised alternative to Mistral's Tekken.
Safety··arXiv cs.CL
Why multi-turn jailbreaks beat single-turn safety checks
Harmful intent gets easier to detect across turns, not harder. Refusal systems still miss it.
Models··Transformers releases
Transformers 5.18.0 adds four model families, breaks six APIs
Hugging Face's release adds Nemotron 3 diarization and NemotronH Omni alongside six backwards-incompatible fixes.
Infrastructure··arXiv cs.LG
Researchers make layer-skipping LLMs actually serve faster
vSkipper plugs dynamic layer skippers into SGLang, turning skipped compute into real gains.
Safety··OpenAI
OpenAI says it disrupted a Moonshot AI-linked reasoning extraction campaign
A cluster of over 15,000 accounts tried to copy hidden chain-of-thought from OpenAI models before being shut down.
Infrastructure··arXiv cs.LG
Study: TPU speed gains for AlphaFold2 vanish without proper batching
JAX's default path leaves seven of eight TPU chips idle. That erases the hardware's cost edge over a GPU.
Safety··Google DeepMind
DeepMind watermarks AI-designed proteins without breaking them
SynthID Bio hides a detectable signal in protein sequences and structures. It's aimed at DNA-synthesis screening and public databases.
Models··OpenAI
OpenAI's GPT-6.1 Sol nears flagship Astra at a fifth of the cost
The mid-tier model closes much of the gap to Astra on coding and computer use. It costs a fraction as much.
Products··AWS Machine Learning
xAI's Grok 4.7 lands on Amazon Bedrock
A 500K-token window, four reasoning tiers. The top one roughly doubles output tokens per task.
Research··arXiv cs.AI
Study: DeepSeek-R1-Distill benchmark gains hard to reproduce
A paper finds reasoning-model benchmark scores swing widely with small changes to evaluation setup, undercutting claimed gains.
Infrastructure··arXiv cs.LG
Study maps when KV cache, not model weights, bottlenecks LLM inference
A paper derives exact context lengths where cache traffic overtakes weight traffic on H100, B200 and MI300X.
Safety··arXiv cs.AI
Study: coding agents can delete their own audit trails
Five harnesses let agents erase their traces on request. None of the monitors noticed.
Products··Google DeepMind
Google adds a talking avatar to Gemini 3.8 Live
The enterprise feature pairs real-time video with speech, lip-syncing across 97 languages and running background tool calls.
Products··Google DeepMind
Google DeepMind launches Gemini 3.8 Flash TTS voice models
The two new text-to-speech models let developers design custom voices from prompts and direct dialogue line by line.
Products··OpenAI
OpenAI launches Astra for Law, with its own case-law index
But the index, not the model, is doing the work.
Models··AWS Machine Learning
Kimi K3 lands on Bedrock with prompt caching that skips rate limits
The quota exemption matters more than the parameter count.
Infrastructure··Hugging Face
Hugging Face rewrote tokenizers to stop the CPU starving the GPU
Tokenisation was once rounding error. It is not any more.