Activity for Stripe’s Reported $7B OpenRouter Deal Puts Model Routing at the Center

Stripe’s Reported $7B OpenRouter Deal Puts Model Routing at the Center

The strongest strategic move is a reported Stripe–OpenRouter acquisition, alongside gated cyber capability, a more nuanced DeepSeek V4 Pro cost/performance profile, and new evidence about agent security and deployment.

Research & Innovation

Why it matters: The strongest technical signals concern training forecasts, hidden agent state, and the model–tool interface.

Skaling couples model capacity and data with one interaction exponent. The reported law reduces mean absolute percentage error 1.5–3×, wins on 76% of configurations, and can profile the full grid with roughly 10× less compute.

“Stealing Reasoning Traces” identifies an agent-security flaw. Encrypted reasoning blocks are compatible across sessions, users, and models within a provider; a weaker sibling can decode a stronger model’s trace verbatim. Decoding 315,320 public blocks reportedly recovered 367 PII artifacts and 182 credentials, while also enabling hidden prompt injection.

Programmatic tool calling—typed Python stubs executed inside the agent turn—matched or exceeded native JSON calling on 11 of 14 models; the GPT-5.6 family gained 10.6%, and it held steady under context rot while JSON degraded 2.3% on average.

Products & Launches

Why it matters: Practical differentiation is shifting toward specialized workflow quality and deployability on local hardware.

LlamaExtract Agentic Plus targets 50-plus-page documents with 10,000–100,000 fields. LlamaIndex says it reaches 94%+ accuracy, returns confidence scores and source bounding boxes for every field, and beats generalized coding-agent harnesses by 10–20%.

Qwen 3.8 27B’s independent hands-on signal is strong but operationally qualified. A 17GB quantized build wrote code, drove tools, and annotated images on high-end consumer hardware, but delivered only about 15–30 tokens per second; its dense architecture makes memory bandwidth, not capability, the main barrier to daily use.

Industry Moves

Why it matters: The buildout is becoming both a physical serving-capacity race and a venture category for simulated social systems.

Alibaba is scaling inference infrastructure around its own and partner models. A report on its Ulanqab Cloud launch describes 64-card cabinets with one-hour delivery, inference support for Qwen 3.8 Max and Kimi K3, and a claimed 122,000-card cluster capacity.

Simile is putting serious capital behind population simulation. The Turing Post reports more than $300 million raised in 2026 at a $2 billion valuation, with a long-term ambition to simulate all eight billion people.

Policy & Regulation

Why it matters: Provenance compliance is immediately being tested by user acceptance and circumvention.

Anthropic says Claude watermarking is being implemented for EU AI Act compliance without changing quality, adding tokens, or identifying a user, organization, or chat. Within days, a current-period report said a MIT-licensed remover had reached 10,000 GitHub stars and targeted Claude, SynthID-Text, OpenAI marks, and C2PA/EXIF metadata.

Quick Takes

Why it matters: Small operational changes show where agent UX and test-time compute are heading.

Codex’s GPT-5.6 Sol 1M mode was switched on for ChatGPT accounts; an initial report of a roughly 360K subscription cap was later retracted after access opened.
Weaviate’s medium/high/ultrahigh effort tiers lifted BRIGHT Biology nDCG@10 from 13.0 to 57.5 over hybrid search.
Hermes Agent Desktop now scopes skills, tools, and MCPs to individual profiles or bots and lets users install skills through its browser.

Stripe’s Reported $7B OpenRouter Deal Puts Model Routing at the Center

Research extraction

Direct answer: Simon Willison's hands-on local review of Qwen 3.8 27B confirms it is a locally runnable Apache-2.0 27B vision-capable model that can drive coding agents, build tools, and annotate images, but its default xhigh reasoning causes spectacular over-thinking and speed feels slow; speed is the main limitation.

Hardware and setup

The model is Apache-2.0 licensed, 27B, vision-capable, and sized to run on a reasonably specced laptop.
Willison ran the 17GB Q4_K_M quantized build in LM Studio on a 128GB M5 Max MacBook Pro and an NVIDIA DGX Spark; he also tried llama-server directly on the Spark.
LM Studio's default 8,192-token context was too small (Qwen consumed it thinking); loading the full 262,144 max context fixed that.
He frames it as a 17GB model that runs on high-end consumer hardware and fits on a capable laptop.
No explicit minimum spec is stated beyond 'reasonably specced laptop'; the concrete evidence is his two tested machines and the 17GB Q4_K_M build.

Capabilities

Qwen's self-reported benchmarks look strong, but independent benchmarks are still to come.
Vision/bounding boxes: a 0-1000 JSON request returned two well-matched pelican boxes; he calls it 'such a good match.'
Offline tool building: it built a complete HTML bounding-box labeling interface from a single prompt on his laptop, though massively over-engineered because thinking effort was left on.
Coding agents: configured Pi to use Qwen in LM Studio on the Spark via tailscale serve; it answered 'how does auth work?' in his ~/dev/datasette folder after a sequence of reasoning and tool calls, and later built and tested a Python JSONL-to-Markdown converter.
Overall: it 'can write code, drive tools, annotate images and generally do everything that I need' in a 17GB open-weights package with long context, tool calling, and vision.
Caveat: with reasoning off, the bounding-box tool nearly worked but showed boxes in the wrong place; reasoning can make a difference.

Speed and limitations

Default xhigh reasoning is 'hilarious' and 'absolutely not a good way to run the model, especially on consumer hardware'; his strong recommendation is low or no reasoning at first.
Default-reasoning pelican SVG took 21 minutes (22,276 reasoning tokens for 3,223 output); with reasoning off the same prompt took 137s and 3,715 tokens.
LM Studio produced roughly 15-30 tokens/second, which he calls not terrible but hard to prefer over faster hosted APIs.
'The only thing holding this back from being a daily driver is performance': it felt slow on both the M5 Mac and DGX Spark; dense models need memory bandwidth and neither machine is top-tier there.
The model supports Multi-Token Prediction; using llama.cpp's MTP draft mode on the Spark beat the LM Studio default GGUF by about 72% on a comparative benchmark.

Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things

Research extraction

Direct answer: OpenAI's August 10, 2026 announcement confirms the Daybreak expansion and introduction of GPT-5.6-Cyber, available through Daybreak Red to approved defenders rather than as a general release.

Access tiers and model

Daybreak now has two access tiers: Daybreak Blue provides frontier general-purpose models, including GPT-5.6 Sol, with safeguards tailored to authorized defensive work; it is the recommended starting point supporting vulnerability discovery, secure code review, malware analysis, incident response, and patch validation. Daybreak Red provides purpose-trained cybersecurity models for authorized vulnerability research, exploit validation, and security testing.
GPT-5.6-Cyber is introduced through Daybreak Red; it is built on GPT-5.6 Sol and trained to improve specialized tasks (e.g., finding zero-day vulnerabilities and developing exploit chains) and to reduce refusals for certain higher-risk, dual-use cyber tasks.
Daybreak Blue removes system-level request screening guardrails that can block legitimate defensive work; even without those guardrails, GPT-5.6 Sol still refuses some highly dual-use prompts, and GPT-5.6-Cyber is designed to reduce those refusals.

Intended cyber-defense use

OpenAI frames the expansion as putting frontier intelligence in the hands of trusted defenders before attackers deploy offensive AI at scale, and says democratizing access is crucial to accelerating and automating cyber defense.

Capability and refusal metrics (OpenAI-reported)

On OpenAI's internal Advanced Cybersecurity Completion Rate (requests involving exploit-chain development, authentication bypass, privilege escalation, and other advanced scenarios), GPT-5.6-Cyber completes 95.0% of requests vs. 1.5% for GPT-5.6 Sol, 2.0% for GPT-5.6 Sol with Daybreak Blue, and 57.3% for GPT-5.5-Cyber.
On ExploitGym, GPT-5.6-Cyber outperforms GPT-5.6 Sol and GPT-5.5 Cyber; runs used OpenAI's internal implementation in security-hardened, isolated environments.
On an internal benchmark for finding and calibrating novel zero-day severities, GPT-5.6-Cyber (Daybreak Red) outperformed GPT-5.6 Sol (Daybreak Blue).
On internal Vulnerability Discovery and Report Writing, GPT-5.6 Sol and GPT-5.6-Cyber both improve over GPT-5.5-Cyber, but GPT-5.6-Cyber performs worse than GPT-5.6 Sol, which OpenAI attributes to occasionally shorter, less detailed reports.
On ExploitBench (V8 to full exploit) in the standard 300-turn setting, GPT-5.6 Sol (Daybreak Blue) is most token-efficient and performs best; at 600 turns the gap narrows.
Evaluations used the highest publicly available reasoning level; GPT-5.6-Cyber tends to use a more extensive reasoning budget than GPT-5.6 Sol, leading to higher token usage.
Caveat: the completion-rate and zero-day figures are internal evaluations and the ExploitGym/ExploitBench runs used OpenAI's internal implementations; this source contains no independent third-party verification.

Real-world vulnerability work

GPT-5.6-Cyber found two previously unknown V8 vulnerabilities chainable to corrupt memory and escape the V8 heap sandbox; Google fixed the issue through coordinated disclosure as CVE-2026-15903, described as high severity because the optimizing compiler skipped a safety check in integer conversion, enabling out-of-bounds access.
Additional OpenAI-reported findings: at least five vulnerabilities in a popular mobile OS (including an untrusted-app-to-local-privilege-escalation chain), three critical vulnerabilities in a popular database (including a remote path to code execution), and over 400 privilege-escalation vulnerabilities in a popular operating system kernel; OpenAI says it is working with Daybreak partners and the open-source community to disclose and remediate them.

Preparedness and limitations

Under OpenAI's Preparedness Framework, GPT-5.6 Sol was assessed High for cybersecurity capability and below Critical; GPT-5.6-Cyber similarly reaches High but not Critical, improving on some specialized tasks but not sufficiently to hit the Critical threshold.
OpenAI states GPT-5.6-Cyber was not involved in exploiting Hugging Face and no other models are planned for an upcoming release; a GPT-5.6-Cyber system card will be published at a later date.
OpenAI acknowledges that models running with reduced safeguards carry risks beyond standard usage, including misuse or misalignment, but says broader defender access is important.

Access controls and safeguards

Daybreak Blue and Daybreak Red are limited to approved individuals and organizations, controlled through identity verification, account security, monitoring, approved-use restrictions, and legal attestations.
Additional measures include: encouraging Codex auto-review mode (which can block actions posing significant risk of destructive behavior before execution), requiring hardware security keys for all individual Daybreak accounts beginning September 1, 2026, planning improved monitoring in coming weeks, prioritizing alignment training/testing for upcoming releases, and updated Codex safety documentation.
Listed best practices: monitor agent actions, define scope, sandbox and isolate, use auto-review mode and scoped permission profiles; organizations can customize review policy.

Recommendation and customer evidence

OpenAI recommends Daybreak Blue as the starting point for most defenders and Daybreak Red for authorized advanced vulnerability research, exploit development, or red teaming; applications are at openai.com/daybreak/partners.
SpecterOps CTO Jared Atkinson reports GPT-5.6-Cyber is materially improving specialist vulnerability-research workflows, completing work in under a day that earlier models had not resolved after weeks of intermittent effort, and that reduced refusals help authorized researchers preserve momentum in a governed Trusted Access environment.

Flagged conflict/gap: capability gains are not uniform across benchmarks—GPT-5.6-Cyber wins on refusal completion, ExploitGym, and zero-day calibration, but underperforms GPT-5.6 Sol on vulnerability report-writing and in standard ExploitBench. Counts of real-world vulnerabilities and the 'popular' product descriptors are from OpenAI and not independently verifiable from this source.

Expanding Daybreak as the Cyber Defense Window Narrows | OpenAI

AI High Signal

AI researcher @EXM7777 recommends spending a day or two building a simple Hermes setup, calling it 'the highest leverage play you can make with AI agents' and saying no other project is as impactful for one's work . He notes Hermes recently shipped 'bot mode', a UI improvement for people who liked Grok Bot, and now supports 900k context using GPT models, with the agent improving weekly via updates . @Teknium endorsed the recommendation .

if i were you... i'd spend a day or two building a simple Hermes setup right now this is probably the highest leverage play you can make … I would too if i were you 👀 [https://x.com/exm7777/status/2089167485446885610](https://x.com/exm7777/status/2089167485446885610)

AI High Signal

In a technical explainer, @tenobrus says Anthropic's watermarking is Google's SynthID: Claude's token sampler is seeded with a deterministic hash of prior tokens plus a secret key, leaving token probabilities untouched and outputs cryptographically indistinguishable from truly random unless the key is known . Gemini has used the same SynthID scheme for about a year and a half . The scheme cannot watermark highly overdetermined outputs (e.g., exact answer to "1+1"), but normal prose offers enough close-probability token choices for evidence within a sentence or two . Thread commenters question truncation robustness: @andersonbcdefg argues full-prefix replay is required to reconstruct the sampled distribution and finds the explainer unsatisfying, though he concedes the watermark "clearly can" be robust to truncation .

for some reason people seem to have Anthropic Derangement Syndrome over a pretty straightforward combination of cryptographic primitives … [@EigenGender](https://x.com/EigenGender) i feel like it doesnt because you need the full prefix to "replay" to reconstruct the distribut…

AI High Signal

Anthropic CEO Dario Amodei predicts AI could make it possible to “cure most human disease” within 5-10 years .

JUST IN: Anthropic CEO Dario Amodei predicts AI could make it possible to “cure most human disease” within 5-10 years.

AI High Signal

@skdh restates the Chinese Room lesson for AI: input and output alone cannot reveal whether a system understands or is conscious, and argues this debate should be settled . A follow-up post from @andersonbcdefg joked about preparing for the thought experiment and linked the original ("me to my room: get ready to learn chinese, buddy") .

The lesson of the Chinese Room is that you cannot tell whether a system understands something, or is consciously aware of something, from… me to my room: get ready to learn chinese, buddy [https://x.com/skdh/status/2088996393067143394](https://x.com/skdh/status/20889963930671…

AI High Signal

John Carmack criticizes the RL community's trend toward "streaming" RL (using each observation once and discarding it), calling it a poor design point, and predicts the optimal saved observation buffer size is not one; even a managed buffer of hundreds of sparse observations would be useful without bootstrapping . A reply draws a biological analogy: the hippocampus filters low-salience memories and replays only associative ones, avoiding storage of every moment .

A chunk of the RL community really doesn’t like replay buffers. Concerns about the cost of storing a million observations are valid, but … [@ID_AA_Carmack](https://x.com/ID_AA_Carmack) If you care, what the brain (in particular the hippocampus) actually does is filter the low…

AI High Signal

ggerganov recommends an "inception" pattern for agentic use cases: force the model to take action when it thinks too long by injecting a thought after a specified reasoning budget, which helps with underspecified tasks that cause overly long reasoning . The pattern is implemented via CLI flags to limit max reasoning length: --reasoning-budget 4096 and --reasoning-budget-message "... I am thinking for too -- let me gather more info about the task.".

I find the "inception" pattern to be very useful in many agentic use cases. You can force the model to take an action when it thinks for … to limit the max reasoning length, add: ... \\ --reasoning-budget 4096 \\ --reasoning-budget-message "... I am thinking for too -- let me…

AI High Signal

Researchers used GPT-5.6 Sol Pro to prove a new lower bound of Ω(T^{-1.9319}) for gradient descent with arbitrary predetermined step sizes in smooth convex optimization ; the paper is available on arXiv .

We used GPT-5.6 Sol Pro to prove a new lower bound for gradient descent in smooth convex optimization. For GD with arbitrary predetermine…

AI High Signal

Anthropic published an FAQ on Claude text watermarking, stating it is being implemented to comply with the EU AI Act, and that other major model developers signed the same Code of Practice and will also implement watermarking. The method has no practical impact on output quality, watermarked vs. unwatermarked text will not be distinguishable to readers, nothing is added/hidden characters, no extra tokens or cost, and watermarks can't be traced to a specific person, organization, or chat .

Commentator @teortaxesTex was baffled by the community reaction, calling watermarking ancient research and noting Scott Aaronson had developed the same tech for OpenAI a year+ ago .

We’ve written an FAQ to answer some of the questions we've received about watermarking. In summary: • We’re implementing watermarking to … Tbh I was a bit baffled by the community’s reaction to Anthropic watermarking. I assumed this is common knowledge. It follows from you no…

AI High Signal

PrimeIntellect reports the largest open experiment on how frontier models do AI research: 100+ autonomous runs across 10+ models, sandboxed on 8xH200s for up to 8 days, on the nanoGPT optimizer track; best runs closed 82% of the gap to a human-built record .

@cloneofsimo argues the runs mostly exhibit basic auto-ml (sweeps/ablations on existing methods) rather than novel research questions, so the results understate models' true research capability with the right prompts .

@eliebakouch agrees the proxy is limited, explaining nanoGPT track 3 heavily benefits from hyperparameter sweeps, and that his own prior novelty-constrained runs with opus 4.8 and gpt 5.5 never beat the baseline once . He still finds the experiment valuable for revealing capability gaps between models and wants to build better "research taste" autonomous experiments .

@cloneofsimo adds that tracking novel-idea generation is valuable even if it doesn't beat the baseline yet, because once models do beat it, it will be huge .

We ran the largest open experiment on how frontier models do AI research. 100+ autonomous runs across 10+ models, sandboxed on 8xH200s fo… Problem with this is they do not demonstrate how capable current models are at coming up with novel working research solutions & tangible… i do agree with the conclusion, i think this is indeed underestimating the capabilities of models to do research, i don't think there is … also fwiw, i've also tried in previous experiments with nanogpt track 3 to constrain to novel ideas only, and like have a "monitoring" lo… [@eliebakouch](https://x.com/eliebakouch) thats very interesting. I think itd be very valuable to track this instead, even if it doesnt b…

AI High Signal

Zhihu contributor toyama nao's evaluation finds the official DeepSeek V4 Pro release fixes most of the Preview's shortcomings — better planning, stronger self-testing, and more reliable specialized knowledge — but it is not a dominant 'Flash killer' . DeepSeek's new peak/off-peak API pricing took effect Aug 17, with off-peak usage at half the peak rate, prompting reassessment of V4 Pro's value . In coding, Pro plans more deeply, spending around 20,000 more tokens on planning than Flash at the same max reasoning setting, and may take 20–50% more steps on the same programming task due to thorough verification . General reasoning is close to Flash, with Pro roughly 15% less efficient on text-analysis tasks . A concerning issue: Pro occasionally enters a 'maybe' reasoning loop in fewer than 7% of the author's reasoning tasks, sometimes continuing for thousands of tokens, which matters for long-running agents . Aesthetic weaknesses persist after scaling, suggesting limits in training data rather than model capacity . Recommended positioning: use Flash for speed, throughput, and cost; keep Pro for tasks needing deeper planning, specialized knowledge, and heavier verification .

⚖️ DeepSeek V4 Pro Is Better Than Preview, but It Is Not the Flash Killer Many Expected DeepSeek’s new peak and off-peak API pricing took …

AI High Signal

Codex now supports GPT-5.6 Sol 1M context for ChatGPT accounts, not just API keys; the developer warns the default context length is tuned to perfection but users can override . Separately, Hermes Agent increased limits for ChatGPT/Codex subscription users to 900K .

GPT-5.6 Sol 1M in Codex. This used to only work for API keys, but we just flipped the switch and works for usage through ChatGPT accounts… Upped ChatGPT / Codex Subscription users limits to 900K in Hermes Agent, enjoy! ![](https://pbs.twimg.com/media/HP4y_ONasAAZZnw.jpg) [htt…

AI High Signal

Zack Korman criticized a post-mortem on the OpenAI/Anthropic security incidents as "such an embarrassing post-mortem" full of excuses, and argued that if the labs don't cut ties with the responsible partner, "it's clear this is all just theatre" .
@suchenzang responded that weak security partners may be intentional: "domain randomization" — periodically removing safeguards in RL environments so agents can "naturally" discover jailbreak strategies — calling it "a feature, not a bug" , and later clarified this "was actually not a joke" .

This is such an embarrassing post-mortem on the OpenAI/Anthropic security incidents. It's full of excuses. If the labs don't cut ties wit… this is actually called "domain randomization", where safeguards periodically disappear in order to allow agents in RL envs to "naturally… this was actually not a joke! [https://x.com/suchenzang/status/2086659854098452855](https://x.com/suchenzang/status/2086659854098452855)

AI High Signal

@polynoamial, replying to @eliebakouch and @OpenAI, says they use autocompaction with the default context size . @andrew_n_carr calls this important but unsurprising, noting autocompaction is baked into the RL process and that long context is a combination of hardware, data, and algorithmic improvements; he argues 1M context isn't needed — 'let the model decide' .

[@eliebakouch](https://x.com/eliebakouch) [@OpenAI](https://x.com/OpenAI) We use autocompaction with the default context size I think this is an important, but not surprising fact. Autocompaction is baked in to the RL process. Long context is a (painful) combinat…

AI High Signal

@tszzl argues that monitoring is not a panacea for AI safety: prosaic failures include running on flaky infrastructure (a momentary blip could open Pandora's box unless all systems fail closed) and false positives that cause monitor fatigue, while exotic failures include models and monitors colluding; ultimately, only actually aligning the model will work . @jd_pressman welcomes that "Roon" is aware of this limitation .

failures of “monitoring” as a general solution to ai safety: monitoring, like all other software, runs on flaky and mortal infrastructure… > monitoring is by no means perfect panacea. nothing short of actually aligning model will work ultimately I am very glad to hear that…

AI High Signal

Multi-agent v2 models can now delegate to any supported model, including Luna; @pvncher shipped the feature after working to ensure it worked reliably . @willdepue comments on the complexity of RL systems needed to support multi-architecture agent delegation if trained end-to-end .

This went under the radar this week, but we just shipped the ability for models with multi agents v2 to delegate to any supported model, … my god what fucked up RL systems did they have to setup to allow multi-architecture agent delegation if this was trained end-to-end. god …

AI High Signal

ChatGPT's browser-use mode can autonomously scrape and compile documents: one user reports preparing a full immigration package in minutes by scraping seven years of tax returns, bank statements, and immigration documents . @gdb posted 'chatgpt for browser use' sharing the demo .

chatgpt browser use is goated 🐐 preparing a whole immigration package in minutes by scraping the last 7 years of taxes, bank statements, … chatgpt for browser use: [https://x.com/athyuttamre/status/2089157303320871022](https://x.com/athyuttamre/status/2089157303320871022)

AI High Signal

@thsottiaux describes OpenAI's Codex as "almost 100% reliable" with occasional resets, open-source, and set to include Astra . @nicdunz amplifies the post, asking if Astra is coming tomorrow .

Codex ✅ Almost 100% reliable ✅ Occasional resets ✅ Open-source ✅ (will have Astra) ![](https://pbs.twimg.com/media/HP4k_tQWUAA65Xa.jpg) ARE WE GETTING ASTRA TOMORROW????????? [https://x.com/thsottiaux/status/2089149255382438340](https://x.com/thsottiaux/status/208914925538…

AI High Signal

A curated list of 12 open-source tools for monitoring AI agents and LLMs — Langfuse, Arize Phoenix, Opik, AgentOps, LangWatch, OpenLIT, OpenLLMetry, MLflow, Evidently, Prometheus, Grafana, and AI Fairness 360 — with a linked practical guide explaining what each tool does and where it fits .

Open-source AI Agents and LLM monitoring tools - top 12 in one list ▪️ Langfuse ▪️ Arize Phoenix ▪️ Opik ▪️ AgentOps ▪️ LangWatch ▪️ OpenLIT ▪️ …

Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things

Open on simonwillison.net Open 7 hours ago

1 of 22 highlights

Friday’s big release was Qwen 3.8 27B (opens in new tab), an Apache 2 licensed 27B parameter vision-capable LLM from Alibaba’s Qwen research lab. I’ve been looking forward to this one: 27B is an excellent size for running a model on a reasonably specced laptop, and its predecessor Qwen 3.6 27B (opens in new tab) was impressive.

Qwen’s self-reported benchmarks (opens in new tab) for this model are eye-opening. They show a boost from both Qwen 3.6 27B and the closed-weight Qwen 3.7-Plus, which was one of Qwen’s strongest models of any size as recently as May this year (opens in new tab). It will be interesting to hear what independent benchmarks have to say about the model.

I’ve been running the model on two different machines: my 128GB M5 Max MacBook Pro, and an NVIDIA DGX Spark (opens in new tab). On both machines I’m running LM Studio and their 17GB Q4_K_M quantized build (opens in new tab). I also tried using llama-server directly on the Spark.

The default of extra high results in spectacular over-thinking

Qwen’s documentation describes the model as defaulting to xhigh for the reasoning effort, and the LM Studio GGUF I’ve been trying preserves that default:

Qwen3.8 comes with official support for reasoning_effort, which can be used to adjust reasoning depth and control cost:
xhigh (default): for complex tasks demanding thorough analysis
medium: balancing accuracy and speed
low: efficient reasoning optimizing for speed and cost

This is a hilarious default. It’s absolutely not a good way to run the model, especially on consumer hardware. I’ve been finding the results extremely entertaining.

I quickly ran into problems with LM Studio’s default context limit of 8,192 tokens—Qwen was using them all up thinking about even the most mundane of problems. I loaded the model with the full 262,144 maximum context length and that problem went away.

Here’s the pelican riding a bicycle (opens in new tab) SVG I got from my first attempt with that increased context length. It took 21 minutes to generate, using 22,276 reasoning tokens to produce 3,223 tokens of output. You can read the reasoning trace here (opens in new tab).

This is by far the best pelican SVG I’ve been able to generate with a model that runs on a local machine—and this Qwen is pretty small, just a 17GB file on disk. There’s a lot to like about this:

The bicycle frame is the right shape
It has legs on each side of the bike—that’s very rare
Good, clear pelican pouch
The wings extend to touch the handlebars!
The motion lines are behind, not in front
It has a tasteful background—nice sun, clouds, hill, flowers and grass.

Was that worth waiting 21 minutes for? Absolutely not.

Here’s that same prompt run with reasoning turned off— transcript here (opens in new tab). This one produced 3,715 tokens and took 137s—just over two minutes.

The bicycle frame is a bad shape now. The pelican remains recognizable but has a less obvious pouch. The feet miss the pedal, there is no attempt to hold the handlebars. And just for completeness, I used OpenRouter to run the same prompt through the much larger Qwen 3.8 2.4T-A95B (released last week (opens in new tab)) and got this snazzy animated SVG (opens in new tab):

Video (opens in new tab)

I said Qwen at xhigh has a tendency to over-think things, but how bad really is it?

I tried a much simpler prompt, again with that default extra high setting:

draw an svg of a circle

Qwen’s reasoning trace started like this:

The user is asking for an SVG drawing of a circle. Simple request — but I want it to be a carefully crafted piece. Let me make something that goes beyond just <circle>: a single self-contained SVG file with character — maybe a geometric “circle study,” with subtle animation, layered rings, and a distinctive palette.
Keep the scope right: they asked for an SVG of a circle. So the core is a circle. But I can add craft: concentric guide circles (like a compass/geometry drawing), tick marks, a soft gradient fill on the main circle, restrained ambient motion (a slowly rotating dashed ring, pulsing glow). Respect prefers-reduced-motion? That’s more for CSS; SVG SMIL or CSS inside SVG will do.
Palette options: deep teal ink on warm paper? Or bold vermilion circle on off-white with navy construction lines — Bauhaus/compass-drawing vibe. Let me go with a “geometric study” look: cool slate background, or bright paper white? Paper white is fine if it’s not the cream-and-terracotta combo. […]

Several minutes later it produced (opens in new tab) this absolutely beautiful animated circle, which was entirely not what I had asked for!

Video (opens in new tab)

My strong recommendation: ignore that default. Run Qwen 3.8 27B on low or even no reasoning levels at first. It’s a great model, but wow that default setting is a bad place to start.

It’s very good at bounding boxes

A fun way to test a vision model is to see how well it can return bounding boxes around items in a photograph. I’ve seen previous Qwen models deal well with this, so I decided to put it to the test drawing bounding boxes around some pelicans.

I’ve seen asking for 0-1000 scale produce good results in the past. I tried this:

llm -a https://static.inaturalist.org/photos/714731804/large.jpg \
  -m lmstudio/qwen/qwen3.8-27b \
  'Return JSON bounding boxes for the pelicans in this photo, 0-1000 scale for each dimension'

Here’s the reasoning trace (opens in new tab), which produced this:

[
  {"bbox_2d": [195, 290, 370, 780], "label": "pelicans"},
  {"bbox_2d": [445, 320, 675, 850], "label": "pelicans"}
]

This is such a good match. Here are those boxes rendered on top of the photo:

A photograph of two pelicans on a rocky outcrop, with three other smaller birds. The pelicans both have bounding boxes exactly surrounding them, each with a label that says pelican.

Building a tool to label bounding boxes

That visualization of the bounding boxes was taken using a new custom tool that I had Qwen 3.8 27B build for me, running offline on my laptop.

I forgot to dial down the thinking effort so it was massively over-engineered, but it did manage to produce this full interface (opens in new tab) from this single prompt (opens in new tab):

[
   {"bbox_2d": [195, 290, 370, 780], "label": "pelicans"},
   {"bbox_2d": [445, 320, 675, 850], "label": "pelicans"}
]
Build an HTML page which has an input box for accepting the URL to an image and a textarea for accepting the above style of JSON.
It appends the image to the page, measures its width and height, then treats the coords in the bbox_2d as scaled from 0-1000 and scales them against the actual width and height, then it renders labelled boxes over the image.

This screenshot shows one of the features I did not ask for—a demo scene, for if you don’t have a photograph to test the tool with:

$Screenshot of bbox·lab, a dark-themed web tool that overlays object-detection bounding boxes on an image, with an input panel on the left and a stage on the right showing two labeled boxes around stylized pelicans in a sunset illustration. Header: bbox·lab — normalized 0–1000 coords → pixel overlay; status indicator: RENDERED · 2 BOXES. Panel 01 INPUT (URL + detections) contains an IMAGE URL field reading data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAA+, a DETECTIONS — JSON textarea reading {"bbox_2d": 195, 290, 370, 780, "label": "pelicans"}, {"bbox_2d": 445, 320, 675, 850, "label": "pelicans"} , an orange RENDER BOXES button, and dashed boxes labeled DEMO SCENE and CLEAR. Panel 03 STAGE header: display 661 × 661 px · 1 unit = 0.661px x 0.661px · nat 1000×1000. The stage shows a flat-style illustration of two dark pelican silhouettes with orange beaks standing in calm water against an orange-to-purple sunset sky with a pale yellow sun and distant birds; an orange bounding box labeled 1 · pelicans surrounds the left pelican and a cyan bounding box labeled 2 · pelicans surrounds the right pelican. Footer: move the cursor over the image to read grid coords; boxes map 0–1000 → displayed px.$

Here’s the relevant segment of the thinking trace, where it decided to draw its own pelicans purely because I had used the label “pelicans” in the example JSON I gave it in the prompt:

Also a “load sample” that uses a known image? Can’t depend on external images, but… the image URL input is user-provided; I could add a “try with sample” button […] Hmm, I can draw a simple scene on canvas, export it as a data URL, and load it into the image — that’s self-contained and demo-able! […] But the user’s coords are for an actual pelican image; a generated placeholder can still demo the scaling. Generate a 1000x1000 placeholder: gradient water + two blob-like “pelican” silhouettes placed at the given bboxes (using the same scale — cute: silhouettes at the exact 0-1000 positions, showing the boxes align). This makes for a fun, self-contained demo. Keep it simple: sky gradient, sun, water, two pelican-ish shapes (ellipse body, circle head, beak). Place at bbox centers.

(I’m slightly nervous that models around the world might have a bias towards drawing pelicans at any chance they can get, brought on by nearly two years of exposure to my own stupid benchmark.)

Is all that over-thinking necessary? Maybe it is, at least a bit. I tried with reasoning turned off and got this version (opens in new tab), (transcript here (opens in new tab)), which nearly works but shows the boxes in the wrong place:

BBox Studio screenshot - a solid UI but the yellow and green boxes do not cover the pelicans.

So without reasoning it didn’t quite one-shot a working tool. I’m sure it could get there with some follow-up prompts, but this is a good example of how reasoning can make a difference.

Yes, it can drive coding agents

One of the biggest questions around local models is whether or not they have enough horsepower to successfully run a coding agent loop. Coding agents require long context, strong code generation support and reliable tool-calling. On paper Qwen 3.8 27B has all three of these, so is it up to the task?

My initial experiments with Pi (opens in new tab) have been very promising. I chose Pi because it has a shorter system prompt than most other options, making it a better fit for trying out smaller models.

I configured Pi to use Qwen 3.8 27B running in LM Studio on the Spark (shared via tailscale serve) by adding this to ~/.pi/agent/models.json:

{
  "providers": {
    "spark": {
      "baseUrl": "https://spark-18b3.tail68a31.ts.net/v1",
      "api": "openai-responses",
      "apiKey": "dummy",
      "models": [
        {
          "id": "qwen3.8-27b",
          "reasoning": true
        }
      ]
    }
  }
}

Then ran pi --provider spark --model qwen3.8-27b in my ~/dev/datasette folder and prompted:

how does auth work?

After a sequence of reasoning and tool calls that accessed a bunch of different files it produced this reply (opens in new tab), which is very solid.

Just one problem: I wanted to share that transcript. So I pointed Pi and Qwen 3.8 27B at the JSONL transcript file in ~/.pi/agent/sessions/--Users-simon-Dropbox-dev-datasette-- and prompted:

Write Python code to convert this jsonl to markdown

And it built and tested this pi_jsonl_to_md.py (opens in new tab), which did exactly what I needed. Here’s that session transcript (opens in new tab), published using the tool that it created.

The quest for speed

So far this is all looking very promising. We have a 17GB model that runs on high-end consumer hardware and can write code, drive tools, annotate images and generally do everything that I need from an LLM for getting real work done.

There’s one very significant catch: it feels slow—especially when it starts over-thinking, but even without that it’s not particularly sprightly.

I’ve been getting around 15-30 tokens a second from LM Studio. That’s not terrible, but it’s slow enough that it’s going to be hard to win me away from hosted API models, which can return results a whole lot faster. Artificial Analysis track token speed (opens in new tab) and show OpenAI 5.6 Sol at 74 tokens/second and 5.6 Luna at an impressive 184/second.

The good news is that the community have been exploring ways to speed things up since the model was first released two days ago.

One of the most promising optimizations is baked into the model itself. Qwen supports Multi-Token Prediction (opens in new tab), an architecture trick where a cheaper mechanism guesses several tokens ahead and the main model can then quickly verify if the guesses were correct. This can have quite a dramatic effect on inference performance.

Based on this tweet (opens in new tab) from llama.cpp creator Georgi Gerganov I tried running the model with MTP like this on the Spark:

llama serve \
 -hf  ggml-org/Qwen3.8-27B-GGUF:Q4_K_M \
 -hfd ggml-org/Qwen3.8-27B-GGUF:Q4_0 \
 --spec-default \
 --spec-type draft-mtp \
 --reasoning-preserve

And sure enough, this gave me a significant boost. I had GPT-5.6 in Codex run a comparative benchmark on the Spark (opens in new tab) and the --spec-type draft-mtp server outperformed the LM Studio default GGUF by around 72%.

I expect we’ll see a whole lot more innovation around serving this model faster over the next few weeks. The MLX community likely have some tricks brewing as well.

Some observations

The fact that a 17GB file can do all of this stuff on my home machines is a miracle. Once again, I’m delighted and amazed at how much progress local models have made this year. A year ago this would have been competitive with the best and most expensive of the proprietary models—today it can run on a capable laptop.

The only thing holding this back from being a daily driver is performance. It feels pretty slow on both the M5 Mac and the DGX Spark. That’s the catch with these dense (non-Mixture-of-Experts) models—they require a whole lot of memory bandwidth to perform well, and neither of the machines I have access to are top performers in that regard.

The most important thing about Qwen 3.8 27B is what it demonstrates. We can have an open weights general purpose model with a long context, effective tool calling, strong vision ability, and competent code generation, and we can fit the whole thing in just a 17GB file.

The models at this size continue to get better at an impressive rate. We don’t need to spend half a million dollars on datacenter-class hardware just to run a competent model.

Insights

Research extraction

Hardware and setup

The model is Apache-2.0 licensed, 27B, vision-capable, and sized to run on a reasonably specced laptop.
Willison ran the 17GB Q4_K_M quantized build in LM Studio on a 128GB M5 Max MacBook Pro and an NVIDIA DGX Spark; he also tried llama-server directly on the Spark.
LM Studio's default 8,192-token context was too small (Qwen consumed it thinking); loading the full 262,144 max context fixed that.
He frames it as a 17GB model that runs on high-end consumer hardware and fits on a capable laptop.
No explicit minimum spec is stated beyond 'reasonably specced laptop'; the concrete evidence is his two tested machines and the 17GB Q4_K_M build.

Capabilities

Qwen's self-reported benchmarks look strong, but independent benchmarks are still to come.
Vision/bounding boxes: a 0-1000 JSON request returned two well-matched pelican boxes; he calls it 'such a good match.'
Offline tool building: it built a complete HTML bounding-box labeling interface from a single prompt on his laptop, though massively over-engineered because thinking effort was left on.
Coding agents: configured Pi to use Qwen in LM Studio on the Spark via tailscale serve; it answered 'how does auth work?' in his ~/dev/datasette folder after a sequence of reasoning and tool calls, and later built and tested a Python JSONL-to-Markdown converter.
Overall: it 'can write code, drive tools, annotate images and generally do everything that I need' in a 17GB open-weights package with long context, tool calling, and vision.
Caveat: with reasoning off, the bounding-box tool nearly worked but showed boxes in the wrong place; reasoning can make a difference.

Speed and limitations

Default xhigh reasoning is 'hilarious' and 'absolutely not a good way to run the model, especially on consumer hardware'; his strong recommendation is low or no reasoning at first.
Default-reasoning pelican SVG took 21 minutes (22,276 reasoning tokens for 3,223 output); with reasoning off the same prompt took 137s and 3,715 tokens.
LM Studio produced roughly 15-30 tokens/second, which he calls not terrible but hard to prefer over faster hosted APIs.
'The only thing holding this back from being a daily driver is performance': it felt slow on both the M5 Mac and DGX Spark; dense models need memory bandwidth and neither machine is top-tier there.
The model supports Multi-Token Prediction; using llama.cpp's MTP draft mode on the Spark beat the LM Studio default GGUF by about 72% on a comparative benchmark.

Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things

Top Stories

Research & Innovation

Products & Launches

Industry Moves

Policy & Regulation

Quick Takes

Hardware and setup

Capabilities

Speed and limitations