Last updated: November 5, 2026
Fireworks Ultra-fast is a public-preview serverless mode announced at Forge on November 3, 2026. It targets extreme latency workloads on DeepSeek V4.1 Flash 1017, aiming for 6–8x generation speed (up to 10x on agentic coding) at a premium per-token price, alongside Nexus Desktop and serverless LoRA serving.
Contents
- What changed at Forge
- Serverless modes compared
- How to call Ultra-fast
- Nexus Desktop and plans
- Specialized Model Inference
- Cost tradeoffs and lower-cost options
- FAQ
This guide covers the Forge launches dated November 3, 2026, with facts from the Fireworks Forge announcement, Serverless Modes and pricing docs, Nexus Desktop and Plans pages, Specialized Model Inference overview, and the docs changelog. For the underlying model, see our DeepSeek V4.1 Flash guide.
What changed at Forge
Fireworks’ Forge conference post says the company shipped faster inference options, Nexus Desktop for model choice inside existing tools, serverless serving for trained LoRA adapters, and enterprise controls. The docs changelog entry for November 3, 2026 names Ultra-fast as a new high-speed Serverless mode in Preview. Ultra-fast variants aim for 6–8x generation speed, and up to 10x for agentic coding tasks, without changing model quality. The Forge post reports up to 1,500 tokens per second demonstrated on DeepSeek V4.1 Flash, and a 30-turn agent task falling from 1 minute to 8 seconds at ultra-fast—those figures come from Fireworks’ announcement, not an independent bench.
Related Forge items that matter to power users: Nexus Desktop joining the Nexus CLI; Specialized Model Inference in public preview starting with GLM 5.3 Flash LoRAs; Background Mode described in the Forge post as a half-price runtime for work that can wait; Multi-day Prompt Caching; higher serverless rate limits and Elastic Burst generally available; and free training on Muse Glimmer 30B through November 30, 2026 (SFT, DPO and RL). For another open-weight path on Fireworks, see our Muse Spark Open guide.
Serverless modes compared
Serverless Modes documentation lists four modes. Standard is the default (no service_tier). Priority raises reliability during peak traffic via service_tier: "priority" on chat completions and the Anthropic-compatible messages API. Fast and Ultra-fast use distinct model IDs rather than a service_tier flag.
| Mode | How you select it | Speed goal (per docs) | DeepSeek V4.1 Flash 1017 price (in / cached / out per 1M) |
|---|---|---|---|
| Standard | accounts/fireworks/models/deepseek-v4p1-flash-1017 |
Default shared capacity | $0.30 / $0.006 / $1.20 |
| Priority | Same model + service_tier: "priority" |
Less likely to shed load (503) | $0.375 / $0.0075 / $1.50 |
| Fast (other models) | Router IDs such as .../kimi-k3-fast, .../glm-5p3-fast |
About 100+ tokens/sec | n/a for this DeepSeek slug |
| Ultra-fast (Preview) | accounts/fireworks/routers/deepseek-v4p1-flash-1017-ultrafast |
6–8x generation (up to 10x agentic coding) | $2.40 / $0.048 / $9.60 |
Prices are from the Serverless pricing table. Preview modes can change quickly. Ultra-fast has no Priority column in that table. If you self-host inference instead, compare with our vLLM 0.32 guide.
How to call Ultra-fast
- Create an API key in the Fireworks platform console and export it as
FIREWORKS_API_KEY. - Confirm DeepSeek V4.1 Flash 1017 Standard works with
accounts/fireworks/models/deepseek-v4p1-flash-1017onhttps://api.fireworks.ai/inference/v1/chat/completions. - Switch the
modelfield toaccounts/fireworks/routers/deepseek-v4p1-flash-1017-ultrafast. Do not setservice_tierfor Ultra-fast. - Send the same messages payload you already use. The changelog says quality matches the base model; you pay the Ultra-fast premium rates.
- Watch rate limits and cost. The changelog notes Ultra-fast is priced at a premium and has lower rate limits than the standard model.
- Fall back to Standard or Priority when latency is less critical than dollars per million tokens.
Example shape from the Serverless Modes docs: a JSON body with the Ultra-fast model ID and a user message, plus an Authorization: Bearer header. OpenAI and Anthropic SDKs work when pointed at Fireworks’ compatible base URLs.
Nexus Desktop and plans
Nexus brings open and closed models into tools people already use. Nexus Desktop (macOS) installs from a .dmg or nexus.fireworks.ai/download: drag to Applications, sign in, turn on each tool, then restart that tool so Nexus Auto (auto) appears in the model picker. Supported harnesses include Claude Code, ChatGPT, Codex, Cursor IDE, VS Code, OpenCode, Pi, Copilot and DeepSeek Harness; Claude Desktop has limited support. The Nexus CLI covers scripted setups on macOS and Linux.
Plans are billed at nexus.fireworks.ai, separate from app.fireworks.ai. Pro is $8 per month (1 seat). Team is $10 per seat per month. Both include Nexus Auto plus your own Claude and ChatGPT models; included Fireworks usage pauses when the window is exhausted. Enterprise adds SSO, shared budgets, access controls and direct open-model picks such as glm-latest. Zero Data Retention applies across plans. Platform API keys do not carry a Nexus plan. Pricing details are also summarized on the Nexus Plans page.
Specialized Model Inference
Specialized Model Inference (public preview) loads a LoRA onto shared serverless GPUs. You pay the base model’s per-token rates with no charge to load, keep loaded, or idle. Supported base today: GLM 5.3 Flash (accounts/fireworks/models/glm-5p3-flash), max LoRA rank 64, up to three adapters loaded per account. Load with firectl model load-lora my-lora --serverless --wait (firectl 1.8.15+), then call accounts/<account>/models/my-lora. First request after load can take up to a minute; set client timeouts to at least 300 seconds. For GLM Flash legal variants elsewhere, see our GLM-5.3-FlashX-Legal guide. Training and dataset tooling overlap with our hf CLI guide.
Cost tradeoffs and lower-cost options
Use Ultra-fast when multi-step agents wait on every token and latency compounds. Use Standard or Priority when reliability or cost dominates. For work that can wait, the Forge post describes Background Mode at half price; the documented Batch API bills asynchronous jobs at 50% of serverless pricing on input and output, and Spot Serverless is 30% below Standard for best-effort capacity (no availability guarantee). Spot uses distinct -spot model IDs and returns 503 when capacity is gone—fall back to Standard yourself. Managed agent platforms such as our Mistral Managed Agents guide solve a different problem (hosted agent sessions) rather than raw token latency.
Limitations to plan around: Ultra-fast is Preview and model coverage is currently one DeepSeek slug; Nexus Desktop is macOS-only for now; Specialized Model Inference supports one base model family and three adapter slots; Forge enterprise items such as Sovereign and No Eval are private preview. Always re-check the pricing table before budgeting.
FAQ
What is Fireworks Ultra-fast?
Ultra-fast is a Preview serverless mode on Fireworks that keeps the same model quality while targeting much higher generation speed. As of the November 3, 2026 docs, it is available for DeepSeek V4.1 Flash 1017 via the model ID accounts/fireworks/routers/deepseek-v4p1-flash-1017-ultrafast.
How much does Fireworks Ultra-fast cost?
According to the Serverless pricing page, DeepSeek V4.1 Flash 1017 Ultra-fast (Preview) is $2.40 / $0.048 / $9.60 per 1M tokens for input / cached input / output. That is higher than Standard for the same model ($0.30 / $0.006 / $1.20). Priority tier is not available for Ultra-fast.
How do I enable Ultra-fast on the Fireworks API?
Change only the model string. Call the OpenAI-compatible chat completions endpoint with model accounts/fireworks/routers/deepseek-v4p1-flash-1017-ultrafast and your FIREWORKS_API_KEY. The Serverless Modes docs say Ultra-fast is not a different model—quality stays the same.
What is Fireworks Nexus Desktop?
Nexus Desktop is a macOS app launched with the Forge products that connects tools such as Claude Code, ChatGPT, Codex, Cursor, VS Code, Copilot and DeepSeek Harness to Nexus Auto with one switch per tool. Linux and Windows desktop support is listed as coming soon; use the Nexus CLI there.
How much do Nexus Pro and Team cost?
The Plans docs list Nexus Pro at $8 per month for one seat and Team at $10 per seat per month (up to 200 seats). Enterprise adds SSO, governance and general access to open models on Fireworks serverless. Nexus billing is separate from the Fireworks platform account.
What is Specialized Model Inference on Fireworks?
It is a public preview that serves your LoRA adapters on shared serverless GPUs with per-token billing at the base model's price and no idle serving charge. At launch it supports GLM 5.3 Flash adapters with rank up to 64 and up to three adapters loaded per account on that base model.
Sources
- https://fireworks.ai/blog/fireworks-unveils-new-products-at-forge-to-help-companies-build-specialized-intelligence
- https://docs.fireworks.ai/serverless/serverless-modes
- https://docs.fireworks.ai/serverless/pricing
- https://docs.fireworks.ai/nexus/desktop
- https://docs.fireworks.ai/serverless/specialized-model-inference/overview