Discover. Decide. Execute.

TTG AI is an AI-powered trading and investing platform, built on our own model, MARI,
and our own agent runtime, BRIAN.

🧠 Meet MARI

MARI is the model behind ttg.ai and the MARI desktop app.

Base	Qwen3.8-Flash-Next, fine-tuned by TTG for behavior only, on real member conversations with no personal information
Works from	Live market data, a full trading day of context and live market tools, as one multimodal model
Runs on	BRIAN, our open-source agent runtime · NVIDIA B300 GPUs on RunPod · our own router on Cloudflare

~180B

_{total parameters, sparse mixture-of-experts}

6B

_{parameters active per token}

262,144

_{tokens of context in a single conversation}

100+

_{market tools MARI can call}

Head to head

On long, multi-step agentic work, MARI's base model is level with OpenAI's GPT-6 Luna and GPT-5.6 Terra and ahead of Claude Opus 4.8, with only 6B active parameters, on infrastructure we run.

Agents' Last Exam	MARI (base)	GPT-6 Luna	GPT-5.6 Terra	Claude Opus 4.8
Long-running professional work across 55 fields _score	51.2	50.9	50.4	45.2

Why it matters for us. Agents' Last Exam grades an agent on long, multi-step jobs, scored on how much of the work it gets right. That is what MARI does for every member: pull quotes, order flow, options flow and news, check the member's saved rules and live chart, and turn it into a trade plan. Doing it with 6B active parameters is what makes each answer fast and lets us run the model ourselves.

_{Sources: MARI, the published results for its base model, Qwen3.8-Flash-Next, before fine-tuning. GPT-6 Luna (max effort): OpenAI, Introducing GPT-6 Sol and Luna. GPT-5.6 Terra and Claude Opus 4.8, as reported by OpenAI: OpenAI, GPT-5.6. Each lab ran its own test setup, so treat gaps under about a point as a tie. Agents' Last Exam is reported two ways: the partial-credit score above, which is the metric OpenAI publishes, and a stricter pass@1 that counts only fully solved tasks (MARI: 24.3, in the Benchmarks table below). On that stricter pass rate, MARI's base trails GPT-5.6 Terra (28.0) and Claude Opus 4.8 (27.0), and newer flagships such as Claude Opus 5.5 score higher on both measures.}

What the model is

Tuned on our members' conversations. MARI is Qwen3.8-Flash-Next, fine-tuned with LoRA on 247,600 real member conversations (2,897,437 messages, Nov 21, 2024 – Aug 5, 2026). The adapter was trained on the full-precision weights, then merged in and compressed to the 4-bit format we serve. The training taught behavior, most of it in how MARI reasons and thinks through a problem: how to pitch an answer to a trader's experience level, the asset they trade and their trading style. It learned patterns, not people or market facts: no personal member details went into training, and market data comes from live sources at request time.
Built for what's next. Qwen calls Qwen3.8-Flash-Next an experimental preview of the architecture that will underpin Qwen 4. Building MARI on it now means our serving stack, training pipeline and tooling already run that architecture. When Qwen 4 ships, we plan to re-run the same fine-tune on it and release an updated MARI.
Sparse mixture-of-experts. About 180B parameters in total: a 125B core network, a 51B n-gram lookup table and a 4B draft head. In each of its 48 layers a router picks 10 of 512 specialist experts for every token, plus one shared expert, so only 6B parameters do the work per token. MARI carries the knowledge of a big model and pays the compute of a small one.
Built for long context. Three fast linear-attention layers for every sparse-attention layer. MARI keeps an entire session in view (price history, the member's plan, every tool result) without the cost blowing up as the conversation grows.
Natively multimodal. A built-in vision encoder lets MARI read chart screenshots, images and video alongside text, in the same conversation.
Thinks before it answers. Reasoning is built in, with adjustable effort, and it streams live so you can watch MARI work.
Drafts ahead. A built-in multi-token prediction head guesses the next 3 tokens and the main model checks them in one pass, so answers stream faster.
Compressed without losing its edge. Expert weights run in 4-bit floating point (NVFP4) and attention in 8-bit (FP8). The full checkpoint is about 171 GiB, including the 51B n-gram table, which is offloaded to host memory rather than kept on the GPU.

How it's served

Runs only inside BRIAN. MARI is served exclusively through BRIAN, our self-hosted, MIT-licensed agent runtime (source on GitHub). BRIAN supplies MARI's instructions, member memory, tool calls and reasoning-effort settings, so the behavior members see is ours and the code behind it is public.
Five dedicated NVIDIA B300 GPUs on RunPod's Secure Cloud, one per pod, all running the same pinned, custom-built vLLM image of our own. Each pod runs the full model on a single GPU, with the 51B n-gram lookup table offloaded to host memory. MARI's 4-bit expert weights use NVFP4, the format Blackwell GPUs run natively. Every pod is tested end to end (answers, multi-turn, streamed reasoning, tool calls) before it takes traffic. Next on the roadmap: moving the fleet onto a single 8-GPU node before December 31, 2026.
Our own router on Cloudflare's edge. A Cloudflare Worker we wrote sits in front of the pods. It keeps each conversation on the same pod so its cache stays warm, spills to the next pod under load, and routes around an unhealthy pod automatically.
OpenAI-compatible. Chat Completions and the Responses API, with streamed reasoning and tool calls in both.
Private by design. MARI requests never go to an outside AI provider. Prompts, member data and trading records stay on infrastructure we run.
Tools first. MARI can call more than 100 market tools (quotes, order flow, options flow, news, fundamentals, charting, trade planning). The core set rides with every request and the rest are found by search when needed, so MARI reaches for live data instead of guessing.

What BRIAN adds around MARI

Knows you every turn. Your local time, your saved notes on style and rules, your live chart and your membership come with every message. Memory is kept per member, ranked by relevance each turn, treated as facts rather than instructions, and never crosses to another member. Balances and account numbers are never saved to it.
Long sessions that don't break. When a chat reaches 80% of what MARI can read at once, older parts are summarized and the last 4 exchanges are kept word for word. The summary keeps a section of exact values and is told never to round a number, and old data results are swapped for a note so MARI re-fetches them instead of guessing.
Recovers mid-answer. If the connection drops, MARI picks up exactly where it stopped, up to 3 times, and never re-runs a tool call that already started. A closed browser or a server restart doesn't lose the answer.
Fast follow-ups. The stable part of every request comes first and everything that changes goes at the end, and each conversation stays on the same MARI server, so repeat work is reused.
Your choice of depth. Pick a reasoning level for each message, or let auto choose it from your message and the last few turns.
Guardrails built in. Figures from MARI's calculator and Python step are placed into the answer exactly, a quote with no price comes back as an error instead of a guess, tool results are treated as data and never as instructions, and paid Holy Grail runs start only from your own button.

The numbers

Measured on our production fleet during the US market open, Monday Nov 2, 2026 (9:24–10:27 ET). Real member traffic.

8,616

_{requests in the first hour of trading}

280–334 tok/s

_{single-request speed per pod}

6.3 s

_{median time to a full answer}

99.6%

_{of requests completed without error}

Benchmarks

Published scores for Qwen3.8-Flash-Next, the base model MARI is fine-tuned from, next to three other models. The fine-tune targets behavior, not capability, so these are the scores MARI starts from. Each lab reports its own scores from its own test setup. Bold marks the best score in each row, and ✅ marks where MARI's base beats every other published score in that row.

What it tests	MARI (base)	Claude Opus 4.6 (Max)	DeepSeek-V4.1-Flash-1017	GLM-5.3-Flash
Agents and tools
Real-world tool use _{Toolathlon Verified, pass@1}	73.5	n/a	n/a	n/a
Multimodal tool use _{ClawEval-MM, average}	60.4 ✅	54.7	n/a	n/a
Professional job tasks _JobBench	55.7 ✅	36.6	n/a	n/a
Frontier agentic tasks _{Agents' Last Exam, pass@1}	24.3	n/a	35.4	26.3
Reasoning and following instructions
Instruction following _IFBench	81.3 ✅	62.5	n/a	n/a
Graduate-level science _{GPQA Diamond}	91.7 ✅	91.3	n/a	n/a
Hardest exam questions _{Humanity's Last Exam}	35.9	40.0	42.6	n/a
Coding
Competitive coding _{LiveCodeBench v6}	91.9 ✅	88.8	n/a	n/a
Agentic coding _{SWE-bench Pro*}	62.5 ✅	53.4	n/a	n/a
Agentic coding _{DeepSWE v1.1}	58.7	n/a	77.0	63.4
Repo-level code generation _{NL2Repo-Bench}	48.1	47.6	68.2	n/a
Vision
Scientific chart analysis _{CharXiv (RQ)}	84.6 ✅	66.0	n/a	n/a
Visual math _MathVision	90.6 ✅	65.5	n/a	n/a
Real-world perception _RealWorldQA	88.5 ✅	73.9	n/a	n/a
Long video understanding _LVBench	76.6 ✅	63.0	n/a	n/a
Size
Parameters active per token	6B ✅	undisclosed	8B–16B	18B

_{n/a means that model has no published score on that benchmark. * MARI's SWE-bench Pro score was run on a corrected version of the task set; Claude Opus 4.6's is its official score.}

Spot checks at serving precision. Third-party spot checks of the base model compressed to the same 4-bit format MARI is served in: thinking on, auto-scored with no LLM judge, on samples of each benchmark (MMLU-Pro 200 questions, MATH-500 100, GSM8K 100, BFCL v4 30), so treat each as accurate to a few points:

89.0

_MMLU-Pro

86.0

_MATH-500

97.0

_GSM8K

96.7

_{BFCL v4 tool calling}

Why it's the right model for TTG

It's built to use tools. Every MARI answer can pull live quotes, order flow, options flow and news. At serving precision, the base model scores 96.7 on a 30-question BFCL v4 tool-calling sample, and the base model's published scores are 73.5 on real-world tool use and 60.4 on multimodal tool use, ahead of Claude Opus 4.6's 54.7.
It handles long, multi-step work. Its base model's Agents' Last Exam score (51.2) puts it level with GPT-6 Luna and GPT-5.6 Terra and ahead of Claude Opus 4.8 (45.2). That benchmark grades the same shape of work as a member's trade plan: many steps, many tools, one deliverable.
It does what you ask. Its base model scores 81.3 on instruction following against Claude Opus 4.6's 62.5. A trading plan has to respect the member's exact rules, size and risk.
It's fast because it's small where it counts. Only 6B parameters fire per token, a third of GLM-5.3-Flash's 18B. That is why a single pod writes 280–334 tok/s and the fleet handled 8,616 requests in the first hour of trading.
It's ours. MARI is built on an open-weight model, carries our own fine-tune and runs on dedicated GPUs under our control. No AI vendor can change, throttle or retire it under us.
Where others lead, and whether it matters. (MARI's numbers here are its base model's published scores.)
- Humanity's Last Exam (DeepSeek-V4.1-Flash-1017 42.6, Claude Opus 4.6 40.0, MARI 35.9) tests expert academic knowledge in math, science and the humanities. MARI's answers come from live market data it pulls with tools, not from academic recall, so this gap doesn't reach members.
- DeepSWE and NL2Repo (DeepSeek-V4.1-Flash-1017 77.0 and 68.2, GLM-5.3-Flash 63.4 on DeepSWE) test building and fixing software inside large codebases. Writing software isn't MARI's job, so these don't affect members either.
- Toolathlon Verified, pass@1 (Claude Opus 5.5 and Claude Sonnet 5.5 77.8, MARI 73.5) does matter. It tests real-world tool use across more than 600 tools in 32 applications, which is MARI's core job. MARI is 4.3 points behind both and ahead of DeepSeek-V4-Flash-0731 (70.3).
MARI's strengths line up with the job: using tools, following instructions and long multi-step work, with only 6B active parameters.

📡 Real-time market data MARI works from

MARI doesn't rely on stale training data. As it works, it pulls real-time data through BRIAN from TTG's own data layer, powered by Nasdaq, Databento, Massive (formerly Polygon.io), Benzinga, Unusual Whales and more: close to 300 live data endpoints across every major asset class, which MARI reaches through its tools.

Level 3 order flow (NL3). A continuous order-by-order stream of the US stock and ETF book: every order added, canceled and filled, as it happens. TTG's engine flags institutional-size orders and powers the Flow Monitor, so MARI sees where size is actually stepping in at a price, not just the last trade.

Stocks. Trades, NBBO quotes, snapshots, gainers and losers, intraday and daily bars, SMA, EMA, MACD and RSI, short interest and short volume, float, splits and dividends, full financial statements, SEC filings including 10-K sections and risk factors, the IPO calendar, and market status and holidays.

Options and options flow. Chains, contract snapshots, trades, quotes, bars and technicals, plus flow analytics: unusual-flow alerts, flow by strike and expiry, gamma and greek exposure (including minute-by-minute spot GEX), max pain, IV rank, IV term structure and skew, open-interest changes, net premium and NOPE, and the day's hottest chains.

Off-exchange and tape. Dark-pool prints, lit-market block prints and the price levels where off-exchange volume is building.

Market-wide positioning. Market, sector and ETF options "tide", net flow by expiry, top net-impact tickers, total options volume and cross-asset correlations.

Ownership and insiders. Institutional holdings and activity, insider buys and sells, congressional trades, and ETF holdings, flows and exposure.

Futures. Contract snapshots, quotes, trades and bars, contract and product specs, trading schedules and automatic front-month resolution.

Indices, forex and crypto. Snapshots, quotes and trades, bars, top movers, technicals and currency conversion.

News and analysts. Breaking headlines filterable by ticker, sentiment and impact, AI summaries with a trading takeaway, analyst ratings, price targets, consensus and bull/bear cases, corporate guidance and events, the FDA calendar, and earnings and dividend calendars.

Macro. CPI and PCE, inflation expectations, Treasury yields, fed funds, SOFR and repo rates, labor-market data and the economic calendar.

Prediction markets. Event odds and their price history: what the crowd is pricing for the Fed, CPI, elections and more.

Screeners and scans. Stock and short-interest screeners, TTG's triangle-pattern scan and a support/resistance engine.

Web. Web and news search, and reading any page as clean text.

TTG's own desk. Moderator trade alerts and chat, the day's watchlists and analyst announcements, the member's tier, credits and course progress, and Holy Grail trade-plan runs, including the Level 3 Flow Monitor.

How MARI uses it

Builds a trade plan from the tape up. It checks price and key levels, reads Level 3 and options flow to see where size is positioned, weighs gamma exposure and max pain, then writes entries, stops and targets that fit the member's rules.
Explains a move as it happens. It lines up the headline, the options flow, the dark-pool prints and the sector tide to say why a stock is moving, not just that it is.
Scans for setups. It runs screeners and pattern scans, then checks each hit against the flow before suggesting it.
Keeps the macro in view. It checks the economic calendar, yields and prediction-market odds before an event, so a plan accounts for CPI or a Fed day.
Works inside the member's account. It knows the member's membership, local time, saved notes and live chart, and uses TTG's own alerts and analyst notes as context.

⚡ Real-time tools Market data is real-time: Level 3 order flow, options flow, dark-pool prints, quotes, futures, indices, forex, crypto, news and prediction-market odds. Reference data (filings, ownership, analyst ratings, macro releases) is as current as its publisher's latest release. All of it is wired in as live tools, not stale training data.	🤖 Agentic trading Agents that scan, research and build trade plans, made for traders who want an edge, not a chatbot.
🔓 Open weights We build on open-weight models and serve them on dedicated GPUs we control. Fast, private, and ours to tune.	🛠️ MCP-native Every capability ships as a tool any agent can call.

Interests: automated & algorithmic trading · quantitative finance · market intelligence · autonomous agents · long-context reasoning · tool use · open-weight inference

_{MARI provides market research and education, not investment advice. TTG AI and its owners and employees are not registered broker-dealers or investment advisers. Consult a licensed professional before making trading decisions.}