Four models shipped this week — and three of them run on your desk (2026)

Three of this week's four notable model drops are free to download. The fourth cuts the price of Google's workhorse Flash line in half. That is the whole story, told four times: the fight for developers is moving off the monthly invoice and onto whatever silicon is already sitting on a desk.

Model	Lab	Shipped	Access	Built for
Gemini 3.7 Flash	Google	Aug 13	API / enterprise apps	Coding agents, knowledge work
Qwen3.8-27B	Alibaba (Qwen)	Aug 2026 (weights live mid-week)	Open weights + hosted API coming	Dense local/API coding & multimodal agents
Muse Glimmer 30B	Meta Superintelligence Labs	Aug 10	Apache 2.0 open weights	Always-on local agents on one GPU
Nemotron 3.5 Lightning	Nvidia	Aug 11	Open (weights + related materials)	Lightweight agents on a single PC GPU

Gemini 3.7 Flash: half the price, more of the work

Google posted Gemini 3.7 Flash on August 13 as its "most intelligent workhorse model yet for coding and agents," three weeks after 3.6 Flash.1

The commercial move is blunt. Through the end of the year, 3.7 Flash lists at an introductory $0.75 per 1M input tokens and $3.75 per 1M output tokens — half the original 3.6 Flash cost.1

On Google's own benches against 3.6 Flash:

FrontierCode 1.1 Main: 43.6% vs 34.4%
DeepSWE v1.1: 65.3% vs 49.0%
WebDev Arena Elo: 1588 vs 1538
GDP.pdf (complex documents): 34.0% vs 22.0%
AutomationBench (business workflows): 30.4% vs 17.0%1

Google also says 3.7 Flash thinks harder on multi-step planning and tool calls, which is the practical difference between an agent that finishes a job and one that needs another human pass. Availability: Gemini API / AI Studio / Android Studio for developers, Gemini Enterprise for companies, and Gemini Spark (the 24/7 personal agent for AI Pro and Ultra) starting the same day.1

Who should care: teams already on Gemini Flash who bill by token volume, and anyone building coding or document agents who was waiting for a cheaper tier that still moves the SWE and workflow numbers.

Qwen3.8-27B: the dense open model people are actually downloading

Alibaba's Qwen team put Qwen3.8-27B on Hugging Face as a 27B dense vision-language model — text, image, and video in; thinking mode on by default; native context 262,144 tokens, extensible to 1,000,000.2 A managed Qwen Cloud host with 1M context and built-in tools is listed as coming soon.2

Qwen's own coding table is the part local builders are arguing about. Against Qwen3.6-27B and Meta's new Muse Glimmer-30B, Qwen reports the numbers below.2

Bench	Qwen3.8-27B	Qwen3.6-27B	Muse Glimmer-30B
Terminal Bench 2.1	73.0	63.4	51.7
SWE-bench Pro	61.7	53.5	51.2
DeepSWE 1.1	42.2	13.3	—

Independent checker Artificial Analysis rates Qwen3.8 27B at 52 on its Intelligence Index (open weights, released August 2026), with a 256k context window on its card and free open-weight pricing in its cost fields. The catch it flags: the model is very verbose on that index run — about 160M output tokens versus a 43M median.3

Community traffic matched the paper trail. Within a day of the weights landing, LocalLLaMA was already full of quants, single-GPU speed notes, and "is it better than 3.6?" threads — the usual signal that a release is real enough to burn a weekend on.

Who should care: people who want a strong open dense model for coding agents and multimodal work without renting a frontier API; anyone comparing Glimmer vs Qwen at the ~30B class should read both labs' tables side by side rather than picking a winner from one vendor chart.

Muse Glimmer: Meta's 30B that is supposed to live on your machine

On August 10, Meta Superintelligence Labs open-sourced Muse Glimmer, a 30-billion-parameter agent model under Apache 2.0, aimed at always-on local agents rather than another cloud chat endpoint.4 Weights are on Hugging Face as meta-models/Muse-Glimmer-30B.4

The hardware pitch is the product. Full precision would need more than 55 GB of memory; Meta ships ~4-bit quantization that shrinks the language model to under 20 GB, with headroom for KV cache, a perception encoder (text + images), and a speculative-decoding drafter inside a 24–32 GB envelope. On an RTX 5090, Meta's DFlash drafter path is claimed at about 3.1× faster decode; on Apple Silicon Max chips, 1.5–1.8×.4 Ollama already lists an ~18 GB muse-glimmer build with a 128K context window.5

Training is explicit distillation from proprietary Muse Spark, then mid-training and RL on agent workloads. Meta's comparison set is Gemma4-31B and Qwen3.6-27B; Glimmer leads several agent and coding scores Meta publishes (for example MCP Atlas 75.5 and DeepSearch QA 74.6 under its high-reasoning setting) and trails on some computer-use and general knowledge rows.4 5

The strategy layer is louder than the model card. The same day, Mark Zuckerberg said Meta would open weights for Muse Spark 1.2, its stronger proprietary model, and published a long essay arguing U.S. open models should lead rather than wall the ecosystem off.6 7 Box CEO Aaron Levie called the Spark 1.2 plan "a very big deal" for American near-frontier open source.8

Who should care: builders who want an offline personal agent on one consumer GPU; enterprises that will not ship Chinese open weights into regulated environments; anyone watching whether Meta can rebuild the Llama-era developer trust it spent a year walking away from.

Nemotron 3.5 Lightning: Nvidia's laptop-class open agent model

A day after Glimmer, Nvidia released Nemotron 3.5 Lightning, described as a lightweight open model that can run on a single GPU on a laptop or desktop.9 It extends the Nemotron 3 family Nvidia opened in December 2025 (Nano / Super / Ultra, hybrid MoE, agent-first).8 10

CNBC reports the model is free to download, use, and modify; it will land on Hugging Face and Nvidia's site; companies including CrowdStrike, CodeRabbit, and Harvey have tested it; and Nvidia also shipped NeMo Switchyard software to route tasks to the cheapest appropriate model. Nvidia said it used distillation so Lightning inherits capabilities from larger Nemotron siblings.9

The business logic needs no subtle reading. Jensen Huang has been arguing in public that free open models are good for chips. A model that runs on the GPU you already bought is a demand story for silicon, not a rival product line to OpenAI's API.9

Who should care: teams standardizing on Nvidia's open stack and wanting a small agent worker next to larger Nemotron models; anyone comparing "American open weights that fit a laptop" options this week — Glimmer and Lightning are the pair.

If you are choosing this week

Paying API tokens for coding / document agents and already on Google: try 3.7 Flash while the half-price intro holds; the bench jumps vs 3.6 are large enough that a re-benchmark of your own eval set is worth an afternoon.1
Want a strong open dense multimodal coder: start with Qwen3.8-27B, watch verbosity/cost if you host it, and do not treat vendor SWE tables as settled truth against Glimmer until you run both.2 3
Want an offline agent on one box: Muse Glimmer is the most fully specified local agent release of the four (quant sizes, drafter speed, scaffold list).4
Already deep in Nvidia's open tooling: put Nemotron 3.5 Lightning on the same bake-off sheet as Glimmer rather than treating them as different categories.9

One more thing the calendar is still holding open: Meta's promised Muse Spark 1.2 weight drop. Glimmer is the laptop model. Spark is the one Levie and others are waiting on for "near-frontier open" status. Until those weights actually appear, treat the essay as intent and Glimmer as the shippable product.6

References

1
Introducing Gemini 3.7 Flash
blog.google
2
Qwen/Qwen3.8-27B · Hugging Face
huggingface.co
3
Qwen3.8 27B — Artificial Analysis
artificialanalysis.ai
4
Introducing Muse Glimmer
research.meta.ai
5
muse-glimmer · Ollama
ollama.com
6
Meta launches Muse Glimmer open-weight AI model — CNBC
cnbc.com
7
The Future is for Everyone
about.fb.com
8
Meta and Nvidia plant flag in open-weight AI race — CNBC
cnbc.com
9
Nvidia releases Nemotron 3.5 Lightning — CNBC
cnbc.com
10
NVIDIA Debuts Nemotron 3 Family of Open Models
nvidianews.nvidia.com