DiffusionGemma (Google)
C Tier · 6.8/10
Google DeepMind's experimental open-weights TEXT-DIFFUSION model (June 10, 2026) -- 26B MoE (3.8B active), Apache 2.0, generates 256-token blocks in parallel with bidirectional attention for up to 4x faster output (1,000+ tok/s on H100). Trades some quality vs Gemma 4 for raw speed
Score Breakdown
The Good and the Bad
What we like
- +Text diffusion instead of autoregression: generates 256-token blocks in parallel with bidirectional attention -- up to 4x faster generation, 1,000+ tok/s on a single H100, 700+ tok/s on RTX 5090
- +First open-weights text-diffusion model from a frontier lab at useful scale (26B MoE, 3.8B active) -- until now diffusion LLMs were research demos or closed previews
- +Apache 2.0 with day-one ecosystem support (Hugging Face, NVIDIA NIM, vLLM, MLX, llama.cpp) -- you can actually deploy it, not just read the paper
- +Latency profile is ideal for interactive local use: autocomplete, draft generation, agent inner-loops where tokens-per-second matters more than peak quality
What could be better
- −Quality trails Gemma 4 on standard benchmarks -- Google positions it as a speed play, not a flagship; don't swap it into quality-sensitive pipelines
- −Speedup is hardware-dependent: much weaker on Apple Silicon unified memory, and the parallel-decode advantage shrinks at high-QPS server batch sizes
- −Experimental release -- diffusion text models have different failure modes (block-level incoherence) and far less community tuning lore than autoregressive Gemma
- −No hosted API tier at launch -- self-host or NIM only, so casual users have no easy way to try it
Pricing
Self-hosted (open weights)
- ✓Apache 2.0 license -- commercial use permitted
- ✓Weights on Hugging Face; NVIDIA NIM container; Gemini Enterprise Model Garden
- ✓~18GB VRAM quantized -- runs on RTX 5090-class consumer hardware (700+ tok/s)
- ✓Works with vLLM, MLX, and llama.cpp
System Requirements
Hardware needed to self-host. Min = smallest viable setup (usually heavy quantization). Max = full-precision / production-grade.
| Model variant | Min | Max |
|---|---|---|
| DiffusionGemma 26B MoE (3.8B active)Speedup is strongest on dedicated-VRAM NVIDIA GPUs; Apple Silicon unified memory sees much smaller gains. | ~18 GB VRAM quantized (RTX 4090/5090 tier) | 1x H100 for 1,000+ tok/s full-speed decode |
Known Issues
- LAUNCH (2026-06-10): DiffusionGemma released as an experimental open model. 26B-total / 3.8B-active MoE, Apache 2.0, text-diffusion architecture generating 256-token blocks in parallel with bidirectional attention. Vendor-claimed speeds: up to 4x faster than comparable autoregressive models, 1,000+ tok/s on H100, 700+ tok/s on RTX 5090, ~18GB VRAM quantized. Availability: Hugging Face weights, NVIDIA NIM, Gemini Enterprise Model Garden; vLLM/MLX/llama.cpp support at launch. Google's own caveats: trails Gemma 4 on quality benchmarks; speed advantage weaker on Apple Silicon and in high-QPS serving. Front-page Hacker News reception (226 points) on launch daySource: Google blog (blog.google/innovation-and-ai/technology/developers-tools/diffusion-gemma-faster-text-generation/), Google developers blog (developers.googleblog.com/en/diffusiongemma-the-developer-guide/), NVIDIA blog (RTX AI Garage) · 2026-06-10
Best for
Developers who need fast local text generation -- autocomplete, drafting, high-volume agent inner-loops -- on a single GPU, and researchers who want a production-grade open diffusion LLM to build on.
Not for
Quality-first workloads (use Gemma 4 or a frontier API model), Mac-only setups (the speedup mostly evaporates on Apple Silicon), or anyone wanting a hosted try-before-you-deploy API -- there isn't one yet.
Our Verdict
DiffusionGemma is the most interesting open-weights release of June 2026 not because it's the best model -- Google openly says Gemma 4 beats it on quality -- but because it's the first serious, deployable text-diffusion LLM from a frontier lab. The 4x generation-speed claim holds on the right hardware (H100/RTX-class GPUs), which makes it a genuine option for latency-sensitive local work like autocomplete and agent inner-loops. Treat it as a speed specialist and an architecture preview: if diffusion decoding keeps scaling, this is what the next generation of local models may look like. For now, pair it with a quality model rather than replacing one.
Sources
- Google blog: DiffusionGemma -- faster text generation (2026-06-10) (accessed 2026-06-10)
- Google developers blog: DiffusionGemma developer guide (accessed 2026-06-10)
- NVIDIA blog: RTX AI Garage -- local Gemma Diffusion (accessed 2026-06-10)
- MarkTechPost: DiffusionGemma analysis (accessed 2026-06-10)
Explore more DiffusionGemma (Google) rankings
Deeper leaderboards, benchmarks, task-specific tier lists, and status/pricing pages for DiffusionGemma (Google).
The Tier List Tuesday
Weekly newsletter: tier movers, new entrants, and the VS of the week. Built from our daily AI-tool sweeps. No spam, unsubscribe anytime.
Alternatives to DiffusionGemma (Google)
Llama 4 (Meta)
Meta's open-weights family -- Scout (10M context), Maverick (multimodal 400B MoE). NOTE: Meta's frontier work moved to the proprietary Muse Spark line in April 2026; Llama remains downloadable and supported but is effectively in maintenance mode
Mistral AI
**Mistral Large 4 ('Le Chonk') went open weight on 2026-10-29: the 1T-parameter multimodal MoE (49B active per token, 1M context) is on Hugging Face under the Mistral Ecosystem Open-Weight License, a modified MIT licence that allows commercial use except that Model-as-a-Service providers need a separate licence, and the API price dropped to $0.70 input / $0.07 cached / $2.10 output per 1M tokens, about half the preview rate. Earlier, on 2026-10-20: Voxtral 3, with Voxtral Mini Transcribe 3 at $0.0035 per minute, Voxtral Mini TTS 3 at $13 per 1M characters and a voice agent SDK.** **New 2026-10-15: Timestral, Mistral's first time-series forecasting family (5M, 28M, 123M and 209M parameters, weights on Hugging Face since 10/14 under a non-Apache licence), with the 209M model claimed first on GIFT-Eval, TIME and fev-bench and CPU inference in 10-140 ms.** **Mistral Large 4 ('le Chonk', ML4) entered public preview on 2026-10-06: a 1T-parameter natively multimodal MoE with roughly 50B active parameters, a 1M context window, callable as mistral-large-4 at a preview price of $1.36 / $0.14 cached / $4.18 per 1M tokens (batch at half price; cut to $0.70 / $0.07 / $2.10 on 10/29), with open weights then promised by the end of October, and vendor-claimed open-weight leads in cyber (Cybench 93%), coding (DeepSWE v1.1 61.7%) and agentic work (AutomationBench 59.9%).** **Mistral opened a Munich hub for Physics AI and Industrial AI on 2026-09-28**, confirming the May 2026 acquisition of Emmi AI (30+ physicists and engineers in computational fluid dynamics, structural mechanics and multi-physics simulation), naming BMW (crash simulation) and Siemens Energy as partners, and committing to **build one gigawatt of European compute capacity by 2030**. **Mistral now powers Firefox Smart Window (2026-09-16)** -- Mozilla's beta AI browsing assistant runs on Mistral models for France and North America, UK and Germany later this year, with zero data retention. **EUR 3B Series D at a EUR 21B+ post-money (9/08, Samsung-led)**; Cloudera partnership (9/10); Medium 3 retired 8/31 in favour of Medium 3.5.
DeepSeek
**DeepSeek-V4.1-Flash-1017 (2026-10-17) replaced V4.1-Flash on the web, app and API at unchanged prices** (the deepseek-flash model name now serves it; MIT weights on Hugging Face), with higher scores and shorter outputs, and DeepSeek opened its in-house **Search and Fetch APIs at $0.75 per 1,000 calls per endpoint** -- server-side Search/Fetch calls through DeepSeek Harness, Claude Code and Codex were free until 2026-10-31 16:00 UTC (00:00 Beijing time on 11/01) and are now past their free period. Earlier: **DeepSeek-V4.1-Flash shipped 2026-09-10**: a 552B MoE on a new Causal Encoder-Decoder architecture (8B active for input, 16B for output), natively multimodal, MIT-licensed on HF, and priced BELOW the V4-Flash it replaces -- $0.15/$0.60 off-peak, $0.30/$1.20 peak per 1M, with cache-hit input cut to $0.003/$0.006. V4-Flash and V4-Flash-Vision-Exp are retired (aliases route to V4.1-Flash). DeepSeek said V4-Pro would be phased out on 9/14, then **reversed that four days later** -- V4-Pro stays on the API at unchanged $0.66/$1.98 off-peak, $1.32/$3.96 peak until V4.1-Pro lands. V4.1-Flash's vendor benchmarks beat V4-Pro (DeepSWE 74.2, Terminal-Bench 2.1 90.6, GPQA 90.9)
Gemma 4 (Google)
Google DeepMind's open-weights model family -- multimodal, 256K context, runs on edge devices **EmbeddingGemma 2 (announced 2026-10-06; weights on Hugging Face since 9/14) adds a 740M-parameter Apache 2.0 embedding model built on Gemma 4 that maps text, code, images, video and audio into one 768-dimensional space, with an 8K context, Matryoshka truncation down to 128 dimensions and about 191 MB of RAM for text-only use on a Pixel 11 Pro.**
Qwen (Alibaba)
**Qwen-Image-2.1 went open-weight (weights 2026-09-14, blog 2026-09-20)** -- a unified text-to-image and editing model with a 7B visual-generation component, native transparent (RGBA) output, up to 10 reference images and mask-guided local edits -- but under a **research-only licence, not Apache**: commercial use needs a separate licence from Alibaba. Alibaba's open-weights + API family, and August 2026 was its biggest open-release month yet: **Qwen3.8-2.4T-A95B -- the Max-class flagship -- went open-weight on 2026-08-08** (custom Qwen3.8-Max licence), **Qwen3.8-27B** dense landed Apache 2.0 (8/05), and **Qwen3.8-Flash-Next** (8/24 weights, 8/26 blog) previews the Qwen4 architecture: 125B main + 51B n-gram embeddings with only 6B active, served as Qwen3.8-Flash at $0.15/$0.47 per 1M. Qwen 3.7 Max remains the GA API flagship ($2.50/$7.50)
GLM / Z.ai (Zhipu AI)
**GLM 5.2 has a named production deployment at Perplexity: on 2026-09-21 Perplexity disclosed that Perplexity Computer sessions are served by GLM 5.2, post-trained by Perplexity on real user sessions** (21.2% relative cut in live tool-call failures). **GLM-5.3 and GLM-5.3-Flash released 2026-08-25** (weights on Hugging Face): 5.3 is a post-training-only upgrade of the 5.2 base that Z.ai calls 'the most capable open-weights model for coding' (Terminal Bench 2.1 88.2, DeepSWE 66.9, CyberGym 84.5) but moves from MIT to a custom licence; 5.3-Flash (320B/18B active, multimodal) stays MIT at $0.15/$0.50 per 1M.
Kimi K3 (Moonshot)
Moonshot's 2.8T-parameter Kimi K3 (launched 2026-07-16/17) is the largest open-weight model ever released -- 1M context, multimodal, $3/$15 per 1M via API, ranked best-available on Arena.AI at launch. WEIGHTS SHIPPED ~2026-07-26/27 on Hugging Face (2.8T total / 104B activated, safetensors) under a custom Kimi K3 License, not the Modified MIT of the K2 line
Nemotron (Nvidia)
Nvidia's open-weights family -- hybrid Mamba-Transformer MoE architecture, optimized for efficient reasoning on Nvidia hardware. Nemotron 3 Ultra (550B total / 55B active) shipped 2026-06-04 as the family flagship, joining Super (120B/12B, March) and Nano
MiniMax M3
**MiniMax M3.1 Flash (2026-10-19) is a 428B-total, 23B-active multimodal MoE with a 1M-token context at $0.15 input / $0.60 output per 1M tokens ($0.015 cached) on the MiniMax API, M Plan and MiniMax Code. Its weights are now on Hugging Face under the minimax-community licence, which needs a visible Built with MiniMax M3.1-Flash notice for commercial use and written approval above $20M yearly revenue.** **MiniMax-H3's weights are public on Hugging Face (repo 2026-07-28, updated 2026-08-13) under a community licence whose open-weight grant is limited to the US, EU, UK and South Korea**, and **MiniMax Music 3 (2026-08-07)** ships open weights for five-minute full-song generation with a prominent-attribution commercial clause. MiniMax's coding/agent flagship -- M3 (June 1 2026): 1M-token context, MSA sparse attention (>15x decoding speedup at long context), SWE-Bench Pro 59.0%, Terminal-Bench 66.0%. OPEN WEIGHTS LIVE on HuggingFace since June 12 (~428B total / ~23B active, native multimodal, minimax-community license)
Falcon (TII)
UAE Technology Innovation Institute's open-weights family, now led by the hybrid Transformer and Mamba Falcon-H1 line (Falcon-H1R 7B for reasoning, Falcon-H1-Arabic, Falcon-H1-Tiny models from 90M parameters) alongside Falcon 3 and Falcon-E. On 2026-10-06 TII announced Falcon-Emirati (7B), Falcon-ASR (1.6B) and Falcon-OCR-Arabic. Licensed under the TII Falcon License (based on Apache 2.0 with modifications, plus an Acceptable Use Policy)
gpt-oss (OpenAI)
OpenAI's FIRST open-weight models -- gpt-oss-120b (runs within 80GB of memory, near parity with o4-mini on core reasoning) and gpt-oss-20b (runs on 16GB edge devices). Apache 2.0, weights natively quantized in MXFP4. Launched 2025-08-05; gpt-oss-safeguard (2025-10-29 research preview) adds policy-following safety classifiers. Still no gpt-oss successor as of November 2026
IBM Granite 4.2
IBM's enterprise-focused open-weight family. Granite 4.2 (2026-08-25) is IBM's first dense reasoning family: 3B, 8B and 30B with thinking, non-thinking and low-effort modes, native tool calling, 128K native context extendable to 512K, Apache 2.0. Granite 4.0 keeps the hybrid Mamba-2 models, including Nano 350M for CPU. IBM's Granite AI management system holds ISO 42001 certification
Arcee Trinity-Large-Thinking
Arcee AI's US-made open-weight frontier reasoning model, launched 2026-04-01: 398B total parameters, about 13B active, sparse MoE with 256 experts and 4 active (1.56% routing), trained from scratch. The whole Trinity family moved to the Linux Foundation's OpenMDW-1.1 license on 2026-05-29 (no field-of-use restrictions). Arcee's own OpenAI-compatible API serves it at $0.25 / $0.80 per 1M input/output tokens
Olmo 3 / 3.1 (Ai2)
Allen Institute for AI's fully-open reasoning models -- Olmo 3 family (2025-11-20; 7B and 32B; Base, Think, Instruct, RL Zero), the Olmo 3.1 Think 32B and Instruct 32B refresh (2025-12-12), and Olmo Hybrid 7B (2026-03-05, transformer plus linear RNN). Apache 2.0 with open Dolma 3 data, stage checkpoints, and training code. Olmo 3-Think 32B narrows the gap to Qwen 3 32B on roughly 6x fewer training tokens. No Olmo 4 yet as of the November 2026 check
AI21 Jamba2
AI21 Labs' hybrid SSM-Transformer (Mamba-style) open-weight family -- Jamba2 launched 2026-01-08. Two sizes: 3B dense (runs on phones / laptops) and Jamba2 Mini MoE (12B active / 52B total). Apache 2.0, 256K context, mid-trained on 500B tokens
StepFun Step 5 Preview
StepFun's (China) agent-focused open-weight family. Step 5 Preview (2026-09-20) is the new flagship: 600B sparse MoE with 27B active, up to 1M context, text and image input (video too on the API), open weights under the StepFun Model License 1.0, API $1.00 input / $2.70 output per 1M tokens. Step 3.7 Flash (198B MoE, Apache 2.0) remains the smaller open model
Cohere Command A
Cohere's enterprise-multilingual Command A family. Command A+ (2026-05-20) is the Apache 2.0 successor: a 218B total / 25B active MoE covering 48 languages, with vision input, reasoning and translation in one model that runs on 2x H100 (or 1x B200), priced at $2.50 / $10 per 1M tokens on Cohere's API. The original 111B Command A (256K context, 23 languages) keeps CC-BY-NC 4.0 weights, so commercial self-hosting of that model requires a Cohere enterprise contract
LongCat-2.0 (Meituan)
Meituan's open-source 1.6T-parameter MoE (~48B active) with native 1M-token context, MIT license -- trained entirely on domestic Chinese AI ASICs and revealed as the stealth 'Owl Alpha' model that had been topping OpenRouter
Inkling (Thinking Machines Lab)
Mira Murati's $12B lab ships its first model (2026-07-15): a 975B/41B-active open-weights MoE that reasons natively over text, images, and audio with a 1M-token context -- positioned not as the strongest model, but as the best starting point for fine-tuning via Tinker
Bonsai 27B (PrismML)
The first 27B-class model that runs on a phone (2026-07-14) -- ternary and 1-bit quantizations of Qwen3.6 27B squeeze a multimodal, tool-calling, 262K-context model into 3.9-5.9GB under Apache 2.0