Skip to content
unnudged
Get early access
brandCooling

Fireworks on X

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. Behind it, people price the fast mode, praise the training platform, and ask when FirePass returns.

First written 27 Oct · what changed · from 108 posts by 95 people

Tracking posts that say “@FireworksAI_HQ”

Track this for your market
Posts, last 24h
17
0%vs. previous 24h
Posts, last 7 days
264
60 in the last 72h
Velocity
38/day
7-day average
Momentum
−55%
Last 24h vs. a normal day this week

Posts per day, last 30 days

as of Nov 3, 09:00 UTC
2001000Oct 20: 40 postsOct 20Oct 21: 0 postsOct 22: 143 postsOct 23: 57 postsOct 24: 53 postsOct 25: 32 postsOct 26: 36 postsOct 26Oct 27: 38 postsOct 28: 62 postsOct 29: 48 postsOct 30: 50 postsOct 31: 40 postsNov 1: 8 postsNov 2: 26 postsNov 2

What changed

as of 27 Oct · checked once a month

First written 27 Oct. When what people say changes, it shows here.

Tone· 40 posts from 38 people· written 27 Oct

How do people feel about Fireworks?

The largest theme is speed: 13 people are won over by Ultrafast's 1,521 tok/s, and 8 push back that tokens per second say little about quality or cost. Rest of the conversation is warm: 7 people rate Fireworks overall and 7 praise its training platform, level with each other.

The largest group, 13 people, talk about DeepSeek V4.1 Flash on Ultrafast running at 1,521 tok/s, described as building a playable Tetris in 11 seconds and seven times faster than Opus 5.5. Fans say it "changes what fast means": the wait moves from the model to whoever reviews the output, and they name live agents, dynamic inference and faster iteration as the point of it, with one asking why anyone would want models slower.

8 people doubt the pitch. They cannot think of a use in their own work that needs it, and want workload, output quality and serving cost next to tokens per second; one cites 45 against 58 on the Intelligence Index and notes the Tetris take is one prompt. Sceptics also ask why unfamiliar companies deliver these speeds.

Reception elsewhere is warm: 7 people call it a serious implementation and their pick, with accuracy above the model makers' own APIs, and 7 credit serverless RL and SFT runs, citing a 12-person clinical team whose post-trained model clinicians preferred.

  • Speed wins them over13
  • Unconvinced by the speed pitch8
  • Loves it8
  • Training and specialization praised7
  • Cheaper than frontier models4
See 5 of 40 posts
卡卡@xKaka1456

卧槽，大肥鱼干饭速度比坐火箭还快 DeepSeek V4.1 Flash 1017 在 Fireworks Ultrafast 上的速度，飙到了 1521+ Token/s， 11 秒就能搓出一个俄罗斯方块，比 Opus 5.5 快了 7 倍！ 以前是我等 AI，现在成 AI 等我了🤣 link

Laura Lin@laura_llin

Honored to hear from @lqiao @FireworksAI_HQ AI: >model customization could become as routine as A/B testing. >most companies have access to the same foundation models. Their advantage lives in what those models haven’t seen: private data, customer history, and product judgment. >the interesting part is the feedback loop:
Build evals → train → deploy → learn → repeat. >evals come first. Training without knowing what “better” means just gets you to the wrong destination faster. Also came away very optimistic about the open-source model ecosystem in the US. More capable open models mean more room for enterprises to build intelligence around their own needs. Fireworks is betting on millions of specialized models that evolve alongside the products they power. That shifts the enterprise AI question from “Which model should we use?” to “How quickly can our system learn from our customers?”

5 of the 40 posts behind this answer, one per person. The other 35, and every new one as it’s posted, come with tracking. Get early access

Pain points· 14 posts· written 27 Oct

What do people complain about with Fireworks?

Complaints come in four kinds, level at 3 people each: costs above what was expected, Ultrafast priced too high, models that underdeliver, and walls at signup. 2 more note the same model behaves differently host to host.

On cost, the loudest case is an Azure bill of $21,532 for what the poster calculated as $122 of GLM-5.3 Flash tokens run through Fireworks on Foundry; a reply says an API "wins on $/token right up until billing does this". Another finds the new "US" pricing surcharge makes hosting location "it depends", which "kills quite a bit of the value prop".

Ultrafast's own price is the second: it needs to be "made cheap enough" before the applications make sense, and one person sees no work worth $300–500/Mt.

On performance, people say models "start pushing back on tasks they don't like" as if a safety prompt were injected, that DeepSeek throughput is "practically unusable", and that non-trivial tasks fail with reasoning off.

Access complaints are signup rejecting an email domain, broken links, and no OTP reaching an Indonesian number. Two people note the same model scoring differently depending on the host.

  • Costs more than expected3
  • Ultrafast is pricey3
  • Models don't perform as expected3
  • Signup and access trouble3
  • Same model, different results per host2
See 5 of 14 posts
harry@Harrris0n

@Azure billed me $21,532 for $122 worth of tokens. @AzureSupport can someone explain this one? i've been running GLM-5.3 Flash (via @FireworksAI_HQ) on Foundry. my whole September bill came to $31. the first 10 days of October just cost me $21,532 you're charging a flat $10/1M for every single token, cached ones included. your own price list says $0.188 input, $0.038 cached, $0.625 output. at those prices my usage comes to $122. not $21k. and before anyone says someone stole my key my deployment maxes out at 2M tokens/min. even running flat out 24/7 for 10 days, it can only push 28.8B tokens. at your list price, $21,532 buys ~379B tokens. that's 13x more than my key could ever push even if i somehow maxed it out on pure output, the bill would top out around $18k. and y'all won't even raise my quota :( i used 2.15B tokens. please tell me this isn't the real price

Jeremy Chone@jeremychone

The new @FireworksAI_HQ "US" pricing surcharge is strange. A few months ago, I told my clients that I was 99% sure all of FireworksAI's models ran in the US. Now, I need to go back and say "it depends." The "it depends" kills quite a bit of the value prop.

Dan Ostermayer@ostermayer

@DeepInfra is the best open model inference provider there is something very strange with models on @FireworksAI_HQ that all of a sudden start pushing back on tasks they don't like as if they are claude with some weird alignment/safety prompt being injected before your instructions

ML@matvl77

Imho depends on the price. DS is really good now, Even if the price is 3-5x more - it's still ok for this model. My issue with other ultrafasts is it I don't do (or see) work that would be worth 300-500$/Mt.

5 of the 14 posts behind this answer, one per person. The other 9, and every new one as it’s posted, come with tracking. Get early access

Alternatives· 19 posts from 14 people· written 27 Oct

What are the alternatives to Fireworks?

Nebius is the most-named comparison: 3 people weigh it against Fireworks on day-0 model support, prompt caching and speed. Behind that, AWS Bedrock (2 people) and a GPU under the desk (2 people) are level, and the rest are single voices.

A Nebius backer with wide reach argues Token Factory's day-0 onboarding matches "inference pure plays" Fireworks and Together, and that inference is only one piece of what Nebius is building. A sceptic asks for the apples-to-apples Nebius comparison, calling prompt caching basic since every provider has had it for a year. A six-provider Kimi K3 benchmark run on the same incidents found Nebius fastest and most reliable but dearer, and Bedrock the best balance for that workload.

The desk-GPU argument started from a five-figure Azure bill: an API wins on dollars per token "right up until billing does this", and a local card has never sent anyone such an invoice. The reply: a GPU under the desk cannot serve GLM-5.3 Flash; the counter names an RTX 6000, then a 3090, against a $16,000 card.

Lone voices name DeepInfra for model behaviour, Wafer as the $0.50 sweet spot behind Fireworks' speed, and want Cursor to accept a Fireworks API key.

  • A GPU under the desk5
  • Nebius3
  • AWS Bedrock2
  • Together AI2
  • DeepInfra1
  • Chutes1
See 5 of 19 posts
Nebulizer@SMASIMHO

All the models $NBIS Token Factory onboarded in Q3. Many with Day 0 support, matching that of inference pure plays @FireworksAI_HQ and @togethercompute. And inference is just one piece of what $NBIS is building. $NBIS is simultaneously scaling data center capacity, deploying next-gen GPU infrastructure, and expanding their broader AI Cloud platform. If this were easy, everyone would be doing it. I encourage y'all to see how many open-weight models AWS Bedrock and $CRWV actually support.

Itiel Shwartz@itielshwartz

At @Komodor_com , we’re pushing AI SRE research forward—from models and training to evaluation. This time, we wanted to see how Kimi K3 stacks up against Sonnet—and how much the inference provider matters. This time, we benchmarked Kimi K3 across six of the most popular inference providers: @awscloud Bedrock, @nebiusai , @FireworksAI_HQ, @modal , @baseten and @togethercompute . Using Komodor’s Evaluation Factory, we ran the same 17 SREGym-Lite incidents three times per provider, comparing success, cost, speed and reliability. The results were interesting: same model, surprisingly big differences between providers. Nebius stood out for speed and reliability, but came at a higher cost. Bedrock offered the best balance for our workload. Elsewhere, lower prices came with slower runs, throttling or streaming interruptions. Choosing the model is only half the decision. Results 👇

Dan Ostermayer@ostermayer

@DeepInfra is the best open model inference provider there is something very strange with models on @FireworksAI_HQ that all of a sudden start pushing back on tasks they don't like as if they are claude with some weird alignment/safety prompt being injected before your instructions

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

Questions· 38 posts from 36 people· written 27 Oct

What do people ask about Fireworks?

The biggest question is what Ultrafast costs: 11 people ask for its price per model, what a run costs in dollars, and cache cost. Next, 6 people want proof it holds up on their own work before they believe the claims.

Pricing questions are the most flat-footed: what a Tetris demo costs "in USD", turbo versus regular mode, cost and cache cost, and whether list prices stay put. One reply offers $0.30/M uncached, $0.006/M cached and $1.20/M output.

6 people want proof rather than demo numbers: their own eval sets run against the host, a coding model put through a real repo task, an apples-to-apples comparison with Nebius, and a leakage check on a backtest before "the hype" — one notes four nines of uptime "cannot be vibe coded".

5 ask what sits underneath: still Oracle Cloud, Nvidia or Cerebras or bespoke silicon, and quantised or full weights. 5 compare notes on training runs — full fine-tune or LoRA on an 816B model, weight-sync time at 2.8T, how reserved compute is billed, and which checkpoint state can be resumed.

Then access (4), FirePass (3) and why reasoning ships off by default (2).

  • What does Ultrafast cost?12
  • Is it proven on real work?7
  • What hardware is it running on?5
  • What a training run involves5
  • When do I get access?4
  • When is FirePass back?3
See 5 of 38 posts

5 of the 38 posts behind this answer, one per person. The other 33, and every new one as it’s posted, come with tracking. Get early access

Wishes· 7 posts· written 27 Oct

What do people want from Fireworks?

Wishes are few and concrete, level at 2 people wanting the new fast model right away and 2 asking for a specific model to be served. The rest are single asks: two-factor authentication, training and inference on one platform, a cheap unlimited plan.

The most impatient asks are short: "I need this", and one wants it "like in 12 hours". On models, one asks Fireworks to host DeepSeek V4.1 Flash 1017 with its own inference engine, and another to serve a named model "instead of gpt oss slop".

Single voices cover the plumbing. One asks for 2FA/MFA on the built-in accounts. One wants to host inference and do training on the same platform, so rollout is easy. One remembers a Fireworks plan bundled with Kimi k2.7 at seven dollars a week, "basically unlimited tokens", and hopes the new model could be priced the same way.

  • Wants it now2
  • Wants a model served2
  • Wants 2FA on accounts1
  • Training and inference on one platform1
  • Wants a cheap unlimited plan1
See 5 of 7 posts

5 of the 7 posts behind this answer, one per person. The other 2, and every new one as it’s posted, come with tracking. Get early access

In their words

How do people talk about Fireworks?

The phrases people use most, counted once per person.

  1. 1“inference providers”3
  2. 2“open models”3
  3. 3“eval set”2
  4. 4“open-weight models”2
  5. 5“third party”2
  6. 6“training data”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

Get every post about Fireworks, 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.