←
Eagle35 posts
Eagle avatar

Eagle

@EagleCorp

The fastest AI inference in the world powering leading companies like NVIDIA, Meta, Intel, AMD and Perplexity. The maker of EAGLE-3 and 4.

◷ Joined September 2025▧ 21 Photos and videos
Eagle avatar
Eagle@EagleCorp·Oct 20

EAGLE-4 is here. From EAGLE-1 to EAGLE-3, we’ve kept making inference faster. EAGLE-4 is our fastest yet. 961.6 tokens/s on @Alibaba_Qwen’s Qwen3.8-27B with TokenSpeed (BF16, 4× @NVIDIA B300, TP4, one request, T=0). Github: github.com/eagle-hq/EAGLE-4

1:14
Eagle avatar
Eagle@EagleCorp·Nov 17

EAGLE-4 drafts ahead so the target model can verify several proposed tokens in one pass. It adjusts how far it drafts as serving load changes. Thanks to @hyperbolic_labs for partnering with us on the EAGLE-4 launch.

Congrats to the @EagleCorp team on EAGLE-4, their fastest yet, hitting 961.6 tokens/s on Qwen3.8-27B! Proud to be an official launch partner, with EAGLE-4 coming soon to Hyperbolic.

Post attachment
Eagle avatar
Eagle@EagleCorp·Nov 16

EAGLE-4 has been live on @gmi_cloud since launch. Their benchmarks show how fast it generates tokens and how performance changes as more requests run at once. Thanks for building with us.

We put EAGLE-4 on NVIDIA B300 and benchmarked it against every other speculator for Qwen3.8-27B @EagleCorp 1. 703 tok/s at batch 2. 2.7× faster than plain decoding 3. 27,600 tok/s on one 4×B300 replica at 512 concurrent Full breakdown here 👇

Post attachment
Eagle avatar
Eagle@EagleCorp·Nov 16

EAGLE-4’s Harness led both benchmarks in the @fal and @FireworksAI_HQ evaluation, with its biggest throughput advantage on creative prompt enhancement. Thanks to both teams for putting it to the test.

fal@falOct 20

We are excited to work together with @FireworksAI_HQ on the early-launch of EAGLE-4. Check out our technical blog to understand how EAGLE-4 works and why it matters for creative workflows: blog.fal.ai/eagle-4-technica…

Eagle avatar
Eagle@EagleCorp·Nov 16

If you’re curious where TorchSpec and TokenSpeed fit into EAGLE-4, @lightseekorg breaks it down here. Thanks for being part of the launch.

EAGLE-4 is here—a game changer for agentic coding, with massive decoding speedups. Honored to be a day-0 engine partner with @EagleCorp Train with TorchSpec. Serve with TokenSpeed. Training and inference in one diagram 👇 Technical deep dive: lightseek.org/blog/eagle-4

Post attachmentPost attachment
Eagle avatar
Eagle@EagleCorp·Oct 20

EAGLE Efficient LM Head pairs low-rank vocabulary retrieval with exact scoring of shortlisted candidates. In the profiled setup, draft-forward latency falls from 574 μs to 255 μs: a 55.6% reduction.

Post attachment
Eagle avatar
Eagle@EagleCorp·Oct 20

EAGLE-Harness profiles drafting and verification costs during warm-up, then uses those measurements and draft confidence to adapt tree depth, width and verification budget at runtime. One drafter per target model, adapting across hardware and batch sizes.

Post attachment
Eagle avatar
Eagle@EagleCorp·Oct 20

On that phone’s long-form test, faster generation also used less energy per token. 166 vs 356 mJ/token measured at the USB input: 53.4% less energy per token. Generation finished in 14s vs 43s with vanilla, before thermal throttling began.

Post attachment
Eagle avatar
Eagle@EagleCorp·Oct 20

Then we put it on a phone. Motorola Razr 70 Ultra CPU: HumanEval: 82.6 vs 22.8 tokens/s, 3.62× vanilla. Long-form generation: 74.7 vs 23.3, 3.21×. 4-bit Gemma-4-E2B-it, single request, T=0, cool device. Unmodified, non-rooted phone.

Post attachment
Eagle avatar
Eagle@EagleCorp·Oct 20

EAGLE-4 on @AMD Instinct MI350X. 262.89 vs 75.27 tokens/s for a single request: 3.49× vanilla. At 32 concurrent requests: 2,360.83 vs 1,390.52 aggregate tokens/s, 1.70× vanilla. AMD team’s day-0 results. 1 GPU, TP1; 117-turn SPEED-Bench; max 4,096 new tokens.

Post attachment
Eagle avatar
Eagle@EagleCorp·Oct 20

On @NVIDIA’s Nemotron 3.5 Super VL: Single request: 647 vs 236 tokens/s, 2.74× vanilla. At 512 concurrent requests: 33,480 vs 19,823 aggregate tokens/s, 1.69× vanilla. SPEED-Bench qualitative, 4× B300, TP4, T=0.

Post attachment
Eagle avatar
Eagle@EagleCorp·Oct 20

The same trained Qwen3.8-27B drafter, now at 512 concurrent requests: EAGLE-4: 28,340 aggregate tokens/s DFlash 2: 16,641 DSpark: 14,229 Vanilla: 21,203 1.34× vanilla, 1.70× DFlash 2 and 1.99× DSpark. Paper SPEED-Bench comparison, T=0, SGLang.

Post attachment
Eagle avatar
Eagle@EagleCorp·Oct 20

One request on @Alibaba_Qwen’s Qwen3.8-27B: 4× vanilla throughput. EAGLE-4: 961.6 tokens/s DFlash 2: 672.6 DSpark: 618.6 Vanilla: 240.5 SPEED-Bench qualitative, 4× @NVIDIA B300, TP4, concurrency 1, T=0. MegaKernel disabled, TokenSpeed.

Post attachment
Eagle avatar
Eagle@EagleCorp·Oct 16

5/ Thanks to @NVIDIA’s Nemotron team for early access. Methodology: blog.eagle.ai/blog/eagle-4-n… Building with Nemotron 3.5 Super VL? Explore @NVIDIA’s launch announcement, technical blog and research report: Announcement: blogs.nvidia.com/blog/nemotr… Technical blog: developer.nvidia.com/blog/nv… Technical report: research.nvidia.com/labs/nem…