Introducing SphotaA state-of-the-art 240M parameter language model

Sphota(sphoṭaḥ)

A meaning-first language model, built around a representation field that connects context with expression. The first release from a different approach to intelligence.

Parameters
240M
One model, under 1B
Overall score
60.14
Average of 14 tasks, highest of 7 models
Compute efficiency
6.1×
Against LFM2-700M
Memory efficiency
4.5×
Against LFM2-700M

01Approach

We introduce Sphota, a 240-million-parameter foundation model built on a simple departure from standard autoregressive design: that meaning should be maintained as an explicit, evolving computational state, rather than left as a transient byproduct of the token sequence that expresses it.

Despite its small footprint and a modest training budget of roughly 2 trillion tokens, Sphota achieves the highest aggregate performance among sub-billion-parameter models in our evaluation suite, reaching 60.14 across 14 benchmarks, outperforming larger and more heavily trained systems including LFM2-700M (55.46), LFM2.5-350M (52.11), MiniCPM4-500M (50.31), and Qwen3-600M (48.03).

02Benchmarks

Sphota’s advantage is sharpest in strict instruction following (79.1 on IFEval), scientific knowledge (94.3 on SciQ), and commonsense reasoning (64.7 on ARC-Easy), More importantly, these results offer an empirical foothold for a larger architectural question: whether separating the computation of meaning from the surface sequence of language can fundamentally shift the capability-to-compute ratio of foundation models.

User-supplied benchmark results for Sphota and six comparison models. Metrics are grouped into instruction following, reasoning, knowledge, science and mathematics. Overall is the average of the task scores.
MODELSIZEOverall ↑Instruction followingReasoningKnowledgeScienceMathematics
IFEval p-sIFEval p-lIFEval i-sIFEval i-lARC-eARC-cPIQAHellaSwagWinograndeMMLUOBQAGPQASciQGSM8K
01
SphotaMannas
240M60.1479.181.785.687.464.737.265.945.460.937.132.826.394.343.4
02
LFM2Liquid AI
700M55.4665.668.675.377.649.238.468.551.755.249.640.625.263.747.0
03
LFM2.5Liquid AI
350M52.1171.775.280.283.045.732.465.743.151.440.736.827.867.68.3
04
MiniCPM4ModelBest / OpenBMB
500M50.3149.555.361.567.047.932.964.444.050.853.333.629.880.633.7
05
Qwen3Alibaba
600M48.0358.262.367.971.547.030.866.242.251.433.235.823.774.57.8
06
Gemma 3Google
270M35.3125.727.438.439.940.627.165.136.554.123.935.222.757.60.1
07
MobileLLM-R1Meta
360M30.3917.718.930.231.934.626.758.733.349.624.532.623.743.00.0

Compared with leading small models from global labs, across 14 tasks.

Every score, grouped by task. The best score in each column is highlighted.

03Efficiency

Sphota builds its representation of the prompt once and generates from that persistent state, so each new token needs fewer operations, and each conversation keeps less state, than in any comparison model. Small scale is an unforgiving testbed: with little capacity to waste, architectural flaws have nowhere to hide. Benchmarks are noisy and small models have hard limits, but organizing a model around a dedicated semantic state lets it do substantially more with fewer parameters, less memory and less compute. Ongoing internal experiments at 1.5 and 2 billion parameters show the same gains carrying forward.

Compute efficiency

Core operations per generated token, against LFM2-700M. Higher is better.

Sphota 240M uses 16% of LFM2-700M’s operations per token.

Memory efficiency

Prompt sessions held per GiB of cache, against LFM2-700M. Higher is better.

Sphota 240M fits 4.5× LFM2-700M’s sessions per GiB of cache.

Training efficiency

Score per unit of training compute, against LFM2-700M. Higher is better.

Sphota 240M gets 15.8× LFM2-700M’s score per unit of compute.

Context length

Native context window, in tokens. Higher holds longer inputs.

Sphota 240M reads 48K tokens, 1.5× every other model’s 32K.

NotesCompute and memory: 512-token prompt, 128 generated tokens, batch 1, BF16, optimal caching for every model; analytical counts of core operations and persistent prompt state, not latency or throughput measurements. Training efficiency is the overall score divided by estimated training compute (parameters × training tokens), relative to LFM2-700M. Training tokens and context length are as stated on each model’s Hugging Face card; Sphota’s are Mannas figures, and its ~2T budget is approximate. MobileLLM-R1 lists 4T of pretraining and 0.2T of mid-training without a per-size total; MiniCPM4’s card adds a 128K extension through YaRN, and its native 32K is shown. Providers count tokens differently, so training budgets are context rather than a measure of training compute.

04Demonstration

Three tasks that often trip up small models: following a chain of relations, finding one detail in a long list, and spotting a structural error. Sphota and three larger models each answer the same prompt, and these are their recorded outputs.

Following the full chain.

Maya is the parent of Leo. Ben is the parent of Alice. George is the parent of Jacob. Alice is the parent of David.

Question

Who is the grandparent of David? Answer with only the name.

ExpectedBen

  1. Sphota240MBenCorrect
  2. LFM2.5350MDavidDoes not match
  3. LFM2700MGeorgeDoes not match
  4. Qwen3600MDavid's grandparent is Alice.Does not match

Recorded CPU outputs. Greedy decoding · BF16 · 8 CPU cores · batch 1 · max 48 new tokens · no web search · no added system prompt.

Native Hugging Face chat templates for the baselines; Qwen thinking off. Sphota uses a Question/Answer wrapper.

05Applications

Sphota is a foundation model, so it is a starting point rather than a finished product: further training adapts it to a specific job. Every application below uses the same Sphota-240M backbone, fine-tuned on only a small number of samples for its task. We have so far fine-tuned it for three kinds of work. The first is games, where it plays chess, the social deduction game Avalon and poker. The second is business decisions, where a 124M-parameter model derived from the same backbone reads a support ticket and its conversation and decides where it goes, how urgent it is and what should happen next. The third is design, where an agent turns a written description into a 3D part. Each application below shows the model at work and, where we have measured it, how it compares with published alternatives.

  1. 01 · Games

    Chess Arena

    A chess-playing model fine-tuned from Sphota.

    Beats ChessLFM
    Plays like Stockfish (Elo)

    The Stockfish strength each model plays like · Sphota scores 77.8% with 1.04 blunders per 100 moves

  2. 02 · Enterprise

    Support ticket routing

    Sphota Decision is a 124M-parameter model derived from Sphota itself. It reads a support ticket and its conversation, then decides its queue, urgency, churn risk, whether a person is needed and the next step, in about 100 ms on a CPU.

    Results on the typed-decisions benchmark
    ModelAccuracy, higher is betterKL, lower is betterBrier, lower is betterECE, lower is better
    Sphota Decision (124M)Fine-tuned · ours (124M)78.20%0.0890.0490.143
    Laya-typed-decisionsFine-tuned · open weights76.60%n/a0.0620.213
    d3 (vLLM-SR)Zero-shot · open weights (27B)74.9%0.4090.1460.049
    Liquid AI d1Zero-shot · API74.2%0.4750.1550.124
    TypeSafe Jev 1.13.0Zero-shot · API72.70%1.4420.1480.144

    Accuracy is agreement with the benchmark’s labels; higher is better. KL, Brier and ECE measure how well the probabilities are calibrated; lower is better.

  3. 03 · Design

    CAD agent

    An agent that creates 3D designs.

    Invalid outputs, %Lower is better
    Median Chamfer distanceLower is better

    Mean Chamfer distance 86.21, against 133.52 for GPT-4o, the next best

  4. 04 · Games

    Avalon Arena

    Five Sphota agents play Avalon, the social deduction game, against each other. Roles are hidden, good or evil, and only speech is shared as they argue over who goes on each of five quests. Each agent says one thing aloud and writes a private note of what it really thinks: as an evil Minion, it notes that it needs to build trust early to secure future failures. In the recorded game, three of five quests failed and the evil side won.

  5. 05 · Games

    Poker

    A fine-tuned Sphota checkpoint reaches 91.03% action accuracy across 11,000 decisions, against 72.14% for GPT 6.1 Sol and 71.71% for Gemini 4 Hydrogen in our evaluations. The original PokerBench paper gives 49.78% for Llama-3-70B, a model about 290 times larger. In match play, the Sphota gameplay checkpoint won two sessions against GPT 6.1 Sol in a five-session heads-up no-limit hold’em match and drew three.

    Results on the PokerBench benchmark
    ModelAction accuracy, higher is betterExact match, higher is better
    Sphota (fine-tuned checkpoint)Fine-tuned · ours (240M)91.03%90.51%
    GPT 6.1 SolAPI72.14%63.13%
    Gemini 4 HydrogenAPI71.71%62.10%
    Llama-3-70BOpen weights (70B) · from the PokerBench paper49.78%39.16%

    Action accuracy and exact match across 11,000 decisions. Sphota, GPT 6.1 Sol and Gemini 4 Hydrogen are Mannas evaluations; Llama-3-70B is the figure from the original PokerBench paper.

Early access

Explore what a different approach to language can make possible.

CONTACT

Get in touch.

A research question, a collaboration, or something you’d like to explore.

Opens a draft in your email app, addressed to hello@mannas.ai. Nothing is sent automatically.