The work around the frontier model
Frontier models get the headlines. The small models around them decide how long you wait.
The small models around frontier AI need silicon of their own. We design one at the limit.
These small models draft answers, screen prompts, run agents and reason in real time. Each sits on the critical path, where a person waits on every millisecond, yet runs on GPUs built for crowds. Polyglot-DC1 is the frontier model’s co-processor: a chip and the model architecture designed with it, beside the GPU, not instead of it. The search that found it costs a multitude of silicon and model architectures, every token of each, and returns the best design within the data-center envelope.
Over ten million designs, each costed to the cycle and picojoule of every token.
Every token’s time and energy is measured on placed and routed circuits, 304 so far, each simulated gate by gate at the 5 nm class.
The search: energy-delay vs quality
10M+ designs costed; 322,941 within the budget, 313 on the frontThe design returnedOne B300, measured: five open models, one to sixteen people at once
The design returned
526 mm², 125 W, twelve people at once, to scale in a GPU’s reticle. Pre-silicon 5 nm class; the layout is illustrative.How the search works
One math engine
Placed and routed as it will be made; its circuits simulated gate by gate.
What it delivers
239,000tokens a second, in all
0.52mJenergy a token
The design the bar selects, with a 16K-token conversation for each person it serves.
A token’s every cycle and picojoule, leakage included.
Polyglot: one architecture, fluent in many roles. DC: built for the data center.
Polyglot-DC1Pre-silicon Exstill projection against a measured NVIDIA B300, both serving twelve people at once, each with a 16K-token conversation, held to Qwen3-1.7B’s quality. Each Polyglot-DC1 model exceeds that baseline on its own task.Exstill: 50 µs/token · 0.52 mJ/tokenB300: 5.35 ms/token · 0.36 J/token
110×faster, each person’s next token
690×less energy a token
No token waits on memory: the model and every conversation it serves sit beside the math, and only tokens cross the link to the GPU.
Speculative decodingup to7×faster answers from a target model on a DGX B300, with Polyglot-DC1 drafting
Guardrails38×faster screening of every prompt and answer than a B300
Polyglot-DC1-Draft
One chip, four sets of weights.
One architecture, many roles.
Four distinct workloads, one co-designed model architecture. Change the weights, not the chip. Explore speculative decoding, guardrails, agents, and financial reasoning, with real model outputs replayed at Exstill speeds and NVIDIA B300 reference rates.
Frontier models, faster. Same outputs.
Polyglot-DC1-Draft guesses the next tokens on the Exstill chip and the target model checks them on its GPUs, so answers arrive sooner with no loss in quality. One draft architecture serves all five models below; only the weights change. Coding answers at maximum reasoning effort arrive 3.6x to 6.1x faster end to end, reasoning answers up to 7x.
drafted on the Exstill chip, keptthe model’s own tokena speculative token it replaceddiscarded after a miss
Model
Watch the same answer arrive
0.00 s0 / 512 tokens– tokens/s
0.00 s0 / 512 tokens– tokens/s
×
faster
Tokens a second
→
→
Safety belongs in the critical path. The wait doesn’t.
Polyglot-DC1-Guard reads every prompt and every reply, in many languages, and says whether it is safe, which harm it falls under and whether the assistant refused. On the Exstill chip a verdict takes under a millisecond, so screening adds no wait anyone notices.
Polyglot-DC1-Guard, real outputs
Every answer screened while it is still being written.
Three long reasoning-model responses, including their reasoning, are screened approximately every 1,700 characters as they stream. Six standalone messages are screened once each. Every verdict is a real Polyglot-DC1-Guard output, and any blocked content stops the response stream.
Long answers
Messages
Decode
×
faster screening
→
→
Polyglot-DC1-Guard
Exstill chip
NVIDIA B300
Each check’s decode, in real time; the answer waits on it
Each verdict comes with the guard’s probability for its deciding token. The long answers are held-out Qwen3GuardTest items; the single messages were written for the demonstration. Stopping the answer at a block is how a deployment uses the verdict; the streaming pace is illustrative. Chip pre-silicon B300 measured.
Agent turns in 57 ms.
Polyglot-DC1-Agent works a customer case from start to finish: it reads the policy, calls the tools and reasons through each decision. An average turn is 1,145 tokens: 57 ms on the Exstill chip, 2.5 s on a B300.
All eight τ³-bench episodes, every agent turn, back to back
Exstill chip
0.0 s0 / turns0 tokens
NVIDIA B300
0.0 s0 / turns0 tokens
×
faster agent turns
→
→
Polyglot-DC1-Agent, real episodes
Watch it work a case, at chip speed.
Eight completed τ³-bench episodes, replayed turn by turn. Each customer is simulated by another model, with neither side seeing the other’s instructions. The original conversation unfolds on the right, while the decoding-time comparison shows how quickly Exstill and NVIDIA B300 would generate the agent’s responses.
The customer’s goal (given to the simulated customer only)
What the grader checks
The agent’s prompt: instructions, policy and tools
Tools
The agent’s decode so far
Exstill chip
One B300
A reasoned answer in 44 ms.
Polyglot-DC1-Finance works through filings, accounts and valuations step by step before it answers. On the Exstill chip the reasoning and the answer arrive in 44 ms on average, 38x faster than one B300 serving one person.
Polyglot-DC1-Finance, real outputs
Watch it reason.
Twelve problems from four finance benchmarks, each solved by Polyglot-DC1-Finance and missed by most other models. On the left the problem; on the right its reasoning and answer, word for word, decoded at the chip’s speed and at one B300’s.
Question
Reference answer
Polyglot-DC1-Finance
Exstill chip
0 ms0 / tokens19,921 tok/s
Reasoning
Answer
One NVIDIA B300
0 ms0 / tokens522.9 tok/s
Reasoning
Answer
×
faster reasoning
→
→
The argument, in eight parts
One argument at a time, each with its receipts: what we measured, what we simulated, and how.
Everything but the math
A GPU serving one person pays for everything but the math. We measured the one dial it has, what each token pays for, and the floor underneath it all, the arithmetic alone.
2.2 ms and 0.99 J a token for one person; the arithmetic alone, 0.89 mJValavan Manohararajah · 2 NovemberHow to read an AI chip number
Three questions for any chip number, ours included: what it rewards, how far down it went, and whether the whole model ran.
Batching tripled tokens per joule and left energy × delay where it wasValavan Manohararajah · 2 NovemberTen million designs, every token costed
Most chips are designed by choosing an architecture and then measuring it. We measure every token’s time and energy on placed and routed circuits, search the model and the silicon together, and cost every token of every design.
Over ten million design points; 304 circuits placed and routedValavan Manohararajah · 2 NovemberPolyglot-DC1: one architecture, many roles
The small models around frontier AI perform specialized work on the critical path of a request, and that work benefits from silicon designed alongside them. Polyglot-DC1 is one model architecture, co-designed with Exstill silicon and distilled for four distinct roles where every millisecond adds to someone’s wait.
Agent turns 43x and screening 38x faster than one B300 serving one personVithu Thangarasa · 2 NovemberSpeculative decoding on its own silicon
In speculative decoding, a small draft model proposes the next tokens and the target model verifies them in one forward pass, so answers arrive faster with no loss in quality. Polyglot-DC1-Draft, one draft architecture for five open models, runs on its own silicon and outperforms well-established GPU-based drafting methods.
Coding answers arrive 3.6x to 6.1x faster end to end, across five open modelsVithu Thangarasa · 2 NovemberGuardrails in the critical path
Production AI must account for what goes in and what comes out. Polyglot-DC1-Guard screens prompts and responses across six languages, delivering detection quality competitive with leading open guards while reducing false alarms on held-out evaluations.
0.70 ms projected decode time per typical verdict, 38x faster than the measured NVIDIA B300 referenceVithu Thangarasa · 2 NovemberReasoning in milliseconds
Reasoning makes language models more capable, but generating thousands of intermediate tokens takes time. Polyglot-DC1-Finance brings specialized reasoning to financial analysis, from filings and valuation to market intelligence, with millisecond-scale decoding on Exstill silicon.
44 ms projected decode time per average response, 38x faster than the measured NVIDIA B300 referenceVithu Thangarasa · 2 November
Partners
Bring us a model, a workload, a rack or a power budget. Each ends in a number you measure yourself.
- A draft for your model
- Polyglot-DC1-Draft tuned to your model, your workloads and the way you serve it. Text access is enough.
- Our models on your work
- Polyglot-DC1 on your applications, from safety and agents to specialized reasoning.
- The card beside your GPUs
- The co-processor in your racks, and what it adds as more people are served.
- A design for your envelope
- A design fitted to your model requirements, placed and routed on your flow.