DiffusionGemma: a parallel-decoding Gemma variant, open-sourced
A team publishing under the "DiffusionGemma" byline — a roster with several familiar Google DeepMind/Gemma contributors, though the paper itself states no institutional affiliation — open-sourced weights for a diffusion-decoding variant of Gemma 4. Instead of generating one token at a time, it refines 256-token blocks in parallel via discrete diffusion, reaching roughly 20 tokens per forward pass and about 1,500 output tokens/second on a single H100, well past autoregressive throughput at a comparable 3.8B activated parameters. It keeps Gemma's thinking mode, multimodal inputs, and long-context support through a two-stage training recipe.
Sources & depth
Mistral ships Agentic Search: multi-step retrieval over one-shot RAG
Mistral released Agentic Search, a five-tool retrieval layer (search, open, navigate, read, grep) that lets a model interrogate documents step by step instead of a single retrieval pass. On table-heavy and financial-filing benchmarks, Mistral reports up to 3x correctness gains and a 45.6-point jump on one benchmark, alongside up to 39.6% lower latency and roughly a third fewer tokens burned per query. It's live now via the Mistral Search Toolkit and built into Studio and Vibe, with a free starter app to try it; no standalone pricing was published.
Sources & depth
CacheRoute: routing requests to where their KV-cache already lives
CacheRoute is a new routing scheme for LLM-serving fleets that assigns each request to the server most likely to already hold its KV-cache prefix, instead of cache-blind load balancing or fixed server affinity. It computes a periodic routing plan that admits high-traffic prefixes to a stable set of servers based on expected load; on Llama-3.3-70B the paper reports lifting cache-hit rate from 64% to 93% and beating the strongest of five baseline routers by 2.3x.
Sources & depth
FleetSieve: profiling only the fleet configurations that matter
Picking tensor-parallel degree and replica count for an LLM serving fleet normally means profiling many configurations, because performance isn't monotonic in TP and the right choice shifts with load. FleetSieve models capacity and tail latency jointly, compares conservative vs. optimistic allocations, and stops measuring once decision uncertainty drops below a threshold. On a 31B model it reached the same fleet-sizing decisions as exhaustive profiling using 6.9% fewer GPU-seconds.
Sources & depth
Liquid AI's DSpark draft models: up to 3.2x faster inference
Liquid AI released DSpark, ~300M-parameter draft models paired with speculative decoding for three of its LFM2.5 language models, claiming up to 3.2x faster inference. The draft model proposes candidate tokens that the larger target model verifies in parallel, amortizing weight-loading cost across more tokens per step; because verification is exact, Liquid says the emitted sequence is identical to greedy decoding on the full model, not an approximation.
Sources & depth
Black Forest Labs launches FLUX Video Upscale, up to native 4K
Black Forest Labs launched FLUX Upscale, a video tool that regenerates footage from as low as 480p up to native 4K while cleaning up blurry faces and texture artifacts, at 1.5x/2x/3x factors. It ships two modes — Precise ($0.07/megapixel-second) for speed and consistency, Creative ($0.10/megapixel-second) for more enhanced detail with a higher chance of altering identity details — available now via BFL's API and web playground.
Sources & depth