Okay guys, if you are renting GPUs for interactive purposes, I don't think anything comes close to @modal. (only now experiencing it and it's🤯) It's built for agents and humans alike. Has no boilerplate about which PyTorch or CUDA package to choose - just launch your jobs
brandRising
Modal on X
The largest group on Modal, 46 of 96 people, posts positive reports, many about setup disappearing rather than one feature. Complaints are about what the bill measures; questions are technical.
First written 27 Oct · what changed · from 103 posts by 96 people
Tracking posts that say “@modal”
Track this for your marketPosts, last 24h
61
+239%vs. previous 24h
Posts, last 7 days
325
113 in the last 72h
Velocity
46/day
7-day average
Momentum
+31%
Last 24h vs. a normal day this week
Posts per day, last 30 days
as of Nov 3, 11:01 UTCWhat changed
as of 27 Oct · checked once a monthFirst written 27 Oct. When what people say changes, it shows here.
Tone· 70 posts from 65 people· written 27 Oct
How do people feel about Modal?
Enthusiasm is the reception: 46 of the section's 65 people post positive reports, most of them praise for the product overall rather than for a feature. The next largest groups praise specific things — elastic GPU capacity (6 people) and the Devin + Modal template (6).
Most of the praise is not tied to a particular feature. People call it the best place to rent GPUs for interactive work and say nothing comes close; newcomers post that they "understand the hype now" after a first job. The line that repeats is that setup disappears: "no boilerplate" about which PyTorch or CUDA package to choose, hard things with infra made easy so people use more AI infra. Users also call it underrated and best in class.
Praise for specific parts splits into small groups. Elastic capacity (6 people) is about bursting from zero to hundreds of on-demand GPUs without reservations, with B200s and B300s reported available and full-day preprocessing reportedly shrinking to minutes. The Devin + Modal template (6 people) is praised for making custom model training and deployment a template link rather than a long setup. Sticky sessions (4 people) get credit for low-latency voice; one developer outside the company writes that its 500ms voice-to-voice is several times faster than most production voice agents they see. Reproducible evals (3 people) are praised for pinning every run to a commit.
- Loves it51
- Elastic GPU capacity praised6
- Devin + Modal template praised6
- Low-latency voice hosting praised4
- Reproducible eval runs praised3
See 5 of 70 postsHide the posts
Our friends at @modal shipped a new Sticky Sessions feature yesterday that's extremely useful for building very low-latency voice agents and other realtime AI applications. Today, you can build (and serve at scale) voice agents that deliver 500ms voice-to-voice response times. But this is 3-4x faster than most of the voice agents we see in production. To be fast, you need a few things: fast models; good orchestration of turn detection and data streaming between models in your agent pipeline; fast WebRTC or telephony client-server network routing; and optimized placement/hosting/routing for your agent and inference containers. Sticky Sessions makes it easy to route a specific user's requests to the same container every time. If you're an infrastructure person, you might describe this as session-aware serverless routing. The Modal team did a lot of work here to get the details right, so things like load balancing and auto-scaling work nicely in conjunction with Sticky Sessions. The video is a demo clip showing a voice agent running on Modal with a 500ms voice-to-voice latency. Ben Shababo, who built this demo, used: - Modal's Qwen 3 ASR service for transcription - Modal's PhoneLLM Mini service as the agent brain - The brand new @GradiumAI Phonon text-to-speech model for the voice - Pipecat as the agent framework and WebRTC routing layer - Modal Sticky Sessions A few really cool things, here. All the models are running co-located on a mixture of GPUs and CPUs: L4, B200, and the Phonon model is running on CPU. Running Phonon on the CPU means that voice audio starts streaming in <50ms and you don't have to manage a GPU cluster for TTS. There are no extra network hops or network overhead in any of the routing. And the agent container uses Modal memory snapshots so a new agent (orchestration code, Phonon model, and the Smart Turn model) are all ready and responding to new requests in under 2 seconds. If you've ever built out voice agent cluster infrastructure on top of, say, Kubernetes, you know how impressive all of the above attributes are! A link to Ben's repo is in the thread below.
Great writeup from @modal on how we’re building general-purpose robotic dexterity. Being able to scale from zero to hundreds of GPUs on demand has made a huge difference in how quickly we can iterate on our world models. Appreciate the Modal team!
custom models went from phd project to a template link
pinning every run to a commit is the part most teams skip, without it you cant tell if the agent changed or the code around it did
5 of the 70 posts behind this answer, one per person. The other 65, and every new one as it’s posted, come with tracking. Get early access
Pain points· 11 posts· written 27 Oct
What do people complain about with Modal?
The bill is the biggest complaint (6 people), mostly about what GPU-hour pricing hides rather than the price itself. Everything else is a handful of single reports: slow starts, limits, and one paying user short of compute.
Billing worries (6 people) are about what the bill measures. One poster argues that pricing post-training by the GPU hour hides the verifier, the eval set and the failed runs thrown away, and that the number that matters is "cost per verified improvement". Another notes that sizing every sandbox for a one-minute test-suite spike gets pricey; a third wants the bill to stop when the task stops. One user who never set a usage limit went past the $30 monthly credit, had the account paused with project data locked inside, and asked for one-time forgiveness.
The rest is scattered. Two people report waits or downtime: a hosted H100 training run that took 21 hours to start, and a passing mention of Modal downtime. Two hit limits — kernel features they need turned off, a 24-hour cap, no RAM snapshots for runtime=vm — and one of them reports GPU allocation, in their agent's words, as a "placement lottery". One paying user asked for help getting compute.
- Billing worries6
- Outages and slow starts2
- Limits and unpredictable GPU placement2
- Hard to get more compute1
See 5 of 11 postsHide the posts
Pricing post-training by the GPU hour hides the real cost. The verifier, the eval set, the failed runs you throw away. The bill that matters is cost per verified improvement, not cost per token trained.
Hey @modal ! I made a rookie mistake by not setting a usage limit on my app and accidentally exceeded my $30 monthly credit. My account (ashish-raj-dev04) is now paused with my crucial project data locked inside. Is there any chance of a one-time invoice forgiveness? 🙏
makes sense, most of a run is waiting on the model and then a test suite or build spikes for a minute. sizing every sandbox for that minute gets pricey
Update: 21 hours later, our @modal-hosted H100 training run is starting to chug away...
ive tried with modal but they simply dont what we need. kernel features we need are off; it can only run for 24h; ram snapshots aren't available for runtime=vm
5 of the 11 posts behind this answer, one per person. The other 6, and every new one as it’s posted, come with tracking. Get early access
Alternatives· 2 posts· written 27 Oct
What are the alternatives to Modal?
Two people weigh Modal against something else, one each: someone who rented GPUs on Salad and Runpod first, and someone who picked E2B sandboxes for easier setup.
The two alternatives are level at one person each, and neither is a head-to-head. The first rented GPUs on Salad and Runpod before moving, and says 5090 availability was the problem there; they describe Modal as cheap, always available, with GPU snapshotting, free volume storage and $30 monthly credits, and call it "cloudflare for inference". The second, in a thread comparing sandbox tools, says they are on E2B because it is easy to setup.
That is all of it. Nobody describes leaving Modal for a specific limit, and no comparison names price, latency or capacity as the deciding difference between the two.
- Runpod1
- E2B1
See the 2 postsHide the posts
@modal is by far the best place to rent GPUs, really cheap, always available, GPU snapshoting, free volume storage and gives you $30 monthly credits! Honesty too good to be true. Its cloudflare for inference. Was using salad & @runpod before but availability of 5090s is an issue
just a pov - I'm E2B and link rn easy to setup
Questions· 19 posts· written 27 Oct
What do people ask about Modal?
No single question dominates. The largest group (5 people) asks about inference internals, then three are level: what people run on it, what it costs, and how evals and RL behave at scale (3, 3 and 3 people).
Inference internals (5 people) is the biggest set: whether cold-start setup eats into an agent's first-request budget, whether the logprob gap between inference engine and trainer on mixed hardware is corrected with importance weights per sample, which component carried sampling stability once quantized samplers joined, and whether selectively quantizing tens of thousands of experts cost quality on the rarely used ones. One person posts their own benchmark of two speculative decoding methods on Modal H200s.
Three questions are level (3, 3 and 3 people): what people actually run on it and whether tool execution or orchestration eats the CPU time; what it costs, from billing only active CPU time to cost per minute, cost per commit and concurrency limits; and how evals and RL fare at scale, from which part breaks first in mixed-hardware scheduling to separating agent regressions from shared-service contention when thousands of runs go in parallel.
Two ask whether the 500ms voice-to-voice figure is first audio byte or full roundtrip, and how much it slips when a tool call lands mid-call. Singles: which sandbox provider someone prefers, GPU credits for a week, and how long a video-data pipeline took with cleaning hours counted.
- Questions about inference internals5
- What do you use Modal for?3
- What does it cost?3
- Questions about evals and RL runs at scale3
- Does the 500ms hold up?2
- Is there an alternative to Modal?1
See 5 of 19 postsHide the posts
curious about cold start latency. when adaption spins up a new agent container, does modal's setup cost eat into the first-request budget?
Rollouts on mixed hardware make the logprob gap between the inference engine and the trainer bigger. Do you correct for it, for example with importance weights per sample, or does it stay small enough to ignore?
what is cost per minute, and what are concurrency limits (eg 200 concurrent)?
At 2,000 parallel runs, I'd be curious how you separate agent regressions from shared-service contention. Running a small serial control alongside the batch, with queue time separated from execution time, could expose failures caused by the eval load itself.
Is that 500ms to first audio byte or full roundtrip? Curious.
5 of the 19 posts behind this answer, one per person. The other 14, and every new one as it’s posted, come with tracking. Get early access
Wishes· 2 posts· written 27 Oct
What do people want from Modal?
Two single requests, level at one person each: serverless, sandboxed Macs, and a way to compare tool traces across commits.
The first wish is serverless, sandboxed Macs. The person asking wants to record demos and promos end to end, sign packages and run MLX models, and thinks Modal "would just win as a service" if it shipped that.
The second is a way for engineers to compare tool traces across commits: when runs last hours, finding the first meaningful difference in document retrieval or tool use would make a change in score easier to diagnose.
Nobody else repeats either request, so these are single asks rather than a pattern of demand.
- Serverless Macs1
- Compare tool traces across commits1
See the 2 postsHide the posts
i feel like @modal would just win as a service unilaterally if they made serverless/sandboxed mac a thing you could e2e, record demos and promos, sign packages, run mlx models, and a ton more
Can engineers compare tool traces across commits? With runs lasting hours, finding the first meaningful difference in document retrieval or tool use would make a score change much easier to diagnose.
In their words
How do people talk about Modal?
The phrases people use most, counted once per person.
- 1“cost per”4
- 2“failed runs”2
- 3“mixed hardware”2
- 4“sticky sessions”2
- 5“voice agents”2
- 6“custom models”2
- 7“model deployment”2
Posts are shown the way X shows them. One of them yours? Here’s why it’s shown, and how to take it off.
Also in Inference and model APIs
- OpenRouter—The conversation is led by 35 people describing their own OpenRouter setup working, with 31 quoting its token and spend figures as the industry's usage signal.388 posts / 7d
- Baseten−25%Approval of how fast Baseten serves open models and how quickly people get running leads the talk, with a free window on its mammoth-1 and elephant-1 decision models drawing recent posts.366 posts / 7d
- Groq−56%Builders credit Groq for what they shipped on it — the largest group, 21 people — with the free tier at 19.320 posts / 7d
- Fireworks−60%The talk is about speed: a DeepSeek V4.1 Flash release on Ultrafast at 1,521 tok/s, with 13 people won over and 8 asking what tokens per second actually buy them.264 posts / 7d
Get every post about Modal, filtered and labelled
This page is the summary. Tracking gives you the stream behind it for any brand, keyword or @mention: each new pain point, request and question as it’s posted, with the spam and promotions already taken out.