LFM2.5-8B-A1B
Compact MoE model for tool calling that abstains instead of hallucinating
About model
LFM2.5-8B-A1B is Liquid AI's compact Mixture-of-Experts model, built on the LFM2 architecture combining MoE routing, grouped-query attention, and gated short convolution blocks. It extends the prior LFM2-8B-A1B release with a 128K-token context window, pre-training scaled from 12T to 38T tokens, a doubled 128K vocabulary that tokenizes non-Latin scripts more efficiently, and large-scale reinforcement learning, and it now produces an explicit reasoning trace before its final answer. The model is built for chaining tool calls and following complex, multi-part instructions, with its strongest results on customer-support-style agent tasks that require sequential API calls. A targeted training stage teaches it to abstain on queries beyond its knowledge rather than answer anyway, reducing hallucination. Available on Together AI.
128K
Roughly 300 pages of text in a single request, up from 32K in the prior release
1B
8.5B total, routed per token, keeping inference cost at small-model levels
38T
Scaled up from 12T in the previous release, followed by large-scale reinforcement learning
- Tool Calling: Handles multi-step tool sequences reliably, with the strongest results on customer-support-style tasks requiring sequential API calls
- Instruction Following: Keeps formatting rules and multi-part instructions across a long conversation
- Hallucination Behavior: Trained to abstain when a query is beyond its knowledge instead of answering anyway
- Production-Ready Infrastructure: 99.9% SLA, available on serverless and dedicated infrastructure
Model | AIME 2025 | GPQA Diamond | HLE | LiveCodeBench | MATH500 | SWE-bench verified |
|---|---|---|---|---|---|---|
LFM2.5-8B-A1B | 42.53 | 88.76 | Related open-source models | Competitor closed-source models | ||
 Claude Opus 4.6 | 90.5% | 34.2% | 78.7% | |||
 OpenAI o3 | 83.3% | 24.9% | 99.2% | 62.3% | ||
OpenAI o1 | 76.8% | 96.4% | 48.9% | |||
GPT-4o | 49.2% | 2.7% | 32.3% | 89.3% | 31.0% |
API usage
Endpoint:
Model card
Architecture Overview:
• LFM2 architecture combining Mixture-of-Experts routing, grouped-query attention, and gated short convolution blocks
• 8.5B total parameters with 1B active per token
• 128K-token context window, expanded from 32K in the prior LFM2-8B-A1B release
• 128K vocabulary, doubled from 65K, improving tokenization efficiency for Hindi, Thai, Vietnamese, Indonesian, and Arabic
• Reasoning-only model that produces an explicit chain of thought before its final answer
Training Methodology:
• Extended pre-training scaled from 12T to 38T tokens versus the prior LFM2-8B-A1B release, followed by large-scale reinforcement learning
• Context extended in stages: a 2T-token midtraining phase to 32K focused on reasoning, math, and tool use, then a 400B-token stage to 128K focused on long-document and long-trajectory data
• A targeted preference optimization stage reduces looping behavior in long reasoning traces
• A reinforcement learning stage using an avg@k-based reward over a diverse knowledge dataset reinforces abstention on queries beyond reliable knowledge
Performance Characteristics:
• Knowledge with hallucination penalty (AA-Omniscience Index, rewards correct answers and penalizes made-up ones): -24.70
• Instruction following (IFEval / IFBench / Multi-IF, adherence to explicit formatting and multi-part instructions): 91.84% / 56.47% / 79.93%
• Math (MATH-500 / AIME 2025 / AIME 2026): 88.76% / 42.53% / 50.00%
• Function calling (BFCLv3 / BFCLv4, correctness of structured tool calls): 64.36% / 48.50%
• Multi-step agent tasks (Tau2-Bench Telecom / Retail, resolving simulated support tickets through sequential tool calls): 88.07% / 39.82%
• Liquid AI reports the model matches larger MoE models such as Gemma 4-26B-A4B on instruction-following benchmarks at a fraction of the active parameter count
Prompting
Together AI API Access:
• Access LFM2.5-8B-A1B via Together AI APIs using the endpoint LiquidAI/LFM2.5-8B-A1B
• Authenticate using your Together AI API key in request headers
• Supports JSON mode and tool calling for structured, multi-step agentic workflows
• The model produces an explicit reasoning trace before its final response
• Available on Together AI serverless and dedicated infrastructure
Applications & use cases
Agentic Tool-Calling Workflows:
• Chain multi-step tool calls for tasks that require sequenced actions and confirmations
• Build telecom and retail-style agent flows that lean on the model's Tau2-Bench strength
• Combine tool calling with JSON mode for structured downstream integrations
Instruction-Heavy Assistants:
• Follow complex, multi-part instructions reliably across conversation turns
• Run assistant workflows that require precise formatting and constraint adherence
• Handle multilingual instructions with tokenization efficient across non-Latin scripts
Cost-Efficient Reasoning at Scale:
• Serve reasoning-backed responses from a compact active parameter count for high-volume workloads
• Run math and structured problem-solving tasks through the Together endpoint
• Scale concurrent agent sessions with the model's compact active footprint
- Model providerLiquid AI
- FeaturesJSON Mode
- DeploymentServerlessMonthly Reserved
- Endpoint
- Parameters8.5B
- Activated parameters1B
- Context length128K
- Input price
$0.03 / 1M tokens
- Output price
$0.12 / 1M tokens
- Input modalitiesText
- Output modalitiesText
- ReleasedMay 27, 2026